Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
3D SPECTROSCOPY IN ASTRONOMY
Volume XVII
Editor in Chief
F. Sánchez, Instituto de Astrofı́sica de Canarias
Edited by
EVENCIO MEDIAVILLA, SANTIAGO ARRIBAS,
M A R T I N M . R O T H , J O R D I C E P A - N O G U É and
F R A N C I S C O S Á N C H E Z
Instituto de Astrofı́sica de Canarias, Tenerife
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore,
São Paulo, Delhi, Dubai, Tokyo
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521895415
List of participants ix
Preface xi
Acknowledgements xii
3 3D spectroscopic instrumentation 87
Matthew A. Bershady
5 Science motivation for integral field spectroscopy and Galactic studies 158
Frank Eisenhauer
vii
Contributors
viii
Participants
ix
x Participants
Jahnke, Knud Max Planck Institute for Astronomy
(Germany)
Jalobeanu, André CNRS Lab. Sciences de l’Image, de
l’Informatique et de la Télédétection (France)
Kehrig, Carolina Instituto de Astrofı́sica de Andalucı́a (Spain)
Kelz, Andreas Astrophysikalisches Institut Postdam
(Germany)
Kjaer, Karina European Southern Observatory (Germany)
Koehler, Ralf Max-Planck-Institut für Extraterrestrische
Physik (Germany)
Kristensen, Egstrom Observatoire de Paris-Meudon (France)
Lagos Lizana, Patricio Andrés Observatorio Nacional ON/MCT (Brazil)
Lara López, Martiza Arlene Instituto de Astrofı́sica de Canarias (Spain)
López Martı́n, Dr. Luis Instituto de Astrofı́sica de Canarias (Spain)
Lorente Espin, Oscar Universidad Politécnica de Cataluña (Spain)
Lyubenova, Mariya European Southern Observatory (Germany)
Maier, Millicent University of Oxford (UK)
Martı́n Gordón, David Instituto de Astrofı́sica de Andalucı́a (Spain)
Moiseev, Alexei Special Astrophysical Observatory (Russia)
Monreal Ibero, Ana Astrophysikalisches Institut Postdam
(Germany)
Mora, Marcelo European Southern Observatory (Germany)
Nowak, Nina Max-Planck-Ins Extraterrestrische Physik
(Germany)
Ocvirk, Pierre University of Central Lancashire (UK)
Parker, Richard John University of Sheffield/ING (UK/Spain)
Pastorini, Guia Universita di Florence (Italy)
Prez Gallego, Jorge University of Florida (USA)
Prez-Montero, Enrique Universidad Autónoma de Madrid (Spain)
Povic, Mirjana Instituto de Astrofı́sica de Canarias (Spain)
Ragaini, Silvia Universita di Padova (Italy)
Ramirez Ballinas, Isidro Universidad Nacional Autónoma Mexico
(Mexico)
Rebaza Castillo, Ovidio Instituto de Astrofı́sica de Andalucı́a (Spain)
Rix, Samantha Anne Isaac Newton Group Telescopes (Spain)
Scheepmaker, Remco Astronomical Institute Utrecht (The
Netherlands)
Smirnova, Aleksandrina Special Astrophysical Observatory (Russia)
Stoklasova, Ivana Observatoire de Lyon (France)
Tamburro, Domenico Max-Planck-Institute for Astronomy
(Germany)
Trachternach, Clemens Ruhr-Universität Bochum (Germany)
Valdivielso Casas, Luisa Instituto de Astrofı́sica de Canarias (Spain)
van Eymeren, Janine Ruhr-Universität Bochum (Germany)
Vidrih, Simon University of Cambridge (UK)
Wang, Wei Peking University (China)
Preface
3D spectroscopy has a relatively short history. Most of the present instrument concepts
were developed in the 1980s and early 1990s. During those pioneering years a great deal of
work was done in optical labs in an attempt to understand how the optical fibres, micro-
lenses and image slicers behave. Only a few groups (often formed by one or two people)
worked on this topic. Communications were not very good, which explains why virtually
all the groups decided to refer to this technique by a different name. So we ended up
with ‘spectral imaging’, ‘bidimensional spectroscopy’, ‘integral field spectroscopy’, ‘two-
dimensional spectroscopy’, ‘3D spectroscopy’, etc.
During those years it was more than doubtful whether this technique was going to
be useful at all. In fact, it looked like a kind of curiosity of limited practical interest
to astronomy. However, in the 1990s the first scientific results were obtained and they
inmediately produced a change of perception.
In the last few years investment in this type of instrumentation has been enormous.
Large telescopes all around the world are now equipped with integral field units. Two
instruments of the future James Webb Space Telescope will also have integral field spec-
troscopic capabilities, etc. Instead of being based in the optical lab trying to characterize
optical fibres or micro-lenses, more effort is dedicated nowadays to refining techniques
for reducing, analysing and interpreting the data obtained with a new generation of 3D
spectrographs. Clearly, we are in a wholly different phase; by attending the lectures and
viewing the posters of our Winter School, it is clear that 3D spectroscopy has truly
arrived on the scene.
This Winter School is particularly timely. It is aimed at instructing a new generation
of astronomers in a new generation of instruments. This will probably be the first and
last school on this topic. In a few years’ time, organizing a similar school would seem like
organizing a school today on ‘imaging’. However, the string of recent scientific conferences
on 3D spectroscopy and the over-subscription to this school would seem to indicate that
this is the right topic at this time.
The Instituto de Astrofı́sica de Canarias (IAC) is specially pleased to be organizing a
Winter School on the topic of 3D spectroscopy because it has contributed to this field
since its inception. More than 15 years ago a small group of researchers at the IAC started
working in this field. They set up from scratch an optical fibre lab at the institute and
developed the first experimental integral field systems for the telescopes on La Palma.
That was just the beginning of the IAC’s contribution to the development and diffusion of
this technique, a contribution that continues today to extend on several fronts, including
the organization of this workshop in collaboration with the European Euro3D network.
The aim of this Winter School is to widen the appeal of integral field spectroscopy
beyond the limited community of experts and to help it become a powerful tool in the
hands of a new generation of astronomers for tackling new (and old) astrophysical prob-
lems. We thank all the participants, lecturers and students for their valuable contributions
to this objective.
The Editors
xi
Acknowledgements
The editors want to express their warmest gratitude to the lecturers for their efforts
in preparing their classes and the chapters of this book. We also wish to thank our
efficient secretaries: Nieves Villoslada and Lourdes Gonzalez, and many staff members
of the IAC: Jess Burgos, Begoña López Betancor, Carmen del Puerto, Ramón Castro,
Terry Mahoney, Miguel Briganti and the technicians of Servicios Informaticos Comunes
del IAC. We acknowledge Ismael M. Delgado for the technical edition of this book for
Cambridge University Press.
Finally, we also wish to thank the Astrophysical Institute Potsdam and the European
Research Training Network Euro3D for their support to the finances, and the organization
and the following institutions for their economical support: The European Commission,
the Spanish Ministry of Education and Science, Iberia, the local governments (cabildos)
of the islands of Tenerife and La Palma, and the Puerto de La Cruz Town Council.
xii
Abbreviations
xiii
xiv Abbreviations
IAU International Astronomical Union
ICA Instituto de Astrofı́sica de Canarias
IDL Interactive Data Language
IFS Integral Field Spectroscopy
IFU Integral Field Unit
IMACS-IFU Inamori Magellan Cassegrain Spectrograph–Integral Field Unit
IRAF Image Reduction and Analysis Facility
ISIS Intermediate-Dispersion Spectroscopic and Imaging System
JWST James Webb Space Telescope
KMOS K-Band Multi-Object Spectrometer
KPNO Kitt Peak National Observatory
MACAO Multi-Application Curvature Adaptive Optics
MCAO Multi-Conjugate Adaptive Optics
MIDAS Mid-Infrared Asteroid Spectroscopy
MIRI Mid-InfraRed Instrument
MOS Multi-Object Imaging Spectrograph
MPE 3D Max Planck Institut für Extraterrestrische Physik 3D
MPFS Multi-Pupil Field Spectrograph
MUSE Multi-Unit Spectroscopic Explorer
NaCo short for NAOS-CONICA (Nasmyth Adaptive Optics System –
Near-Infrared Imager and Spectrograph)
NIC-FPS Near-Infrared Camera and Fabry–Perot Spectrometer
NICMOS Near Infrared Camera and Multi-Object Spectrometer
NIFS Near-Infrared Integral Field Spectrograph
NIRSpec Near-Infrared Spectrograph
NOAO National Optical Astronomy Observatory
NOT Nordic Optical Telescope
OAN-SPM National Astronomical Observatory at San Pedro Martir
OASIS Optically Adaptive System for Imaging Spectroscopy
OHP Observatoire de Haute-Provence
OMM Observatoire du mont Megantic
ORM Observatorio del Roque de los Muchachos
OSIRIS Optical System for Imaging and low/intermediate-Resolution
Integrated Spectroscopy
PACS Photoconductor Array Camera and Spectrograph
PMAS Potsdam Multiaperture Spectrophotometer
PPAK PMAS fiber pack
PSF Point Spread Function
PUMA UNAM (Universidad Nacional Autónoma de Mexico) Scanning
Fabry–Perot Interferometer
PUMILA Near-infrared Spectrograph for the Kinematic Study of the
Interstellar Medium
PYTHEAS Prisme Interferomètre Trames de lentilles pour l’Holometrie, et
l’Endoscopie des Astres et des Sources
QSO quasi-stellar object
Abbreviations xv
RFP Rutgers Fabry–Perot
RSS Robert Stobie Spectograph
RSS Row Stacked Spectra
RTN Research Training Network
SALT Southern African Large Telescope
SAO Smithsonian Astronomical Observatory
SAURON Spectrographic Areal Unit for Research on Optical Nebulae
SHM shared memory server
SILFID Spectrographe Integral Linarisation par Fibres de l’Image
Directe
SINFONI Spectrograph for Integral Field Observations in the Near
Infrared
SIS Stabilized Imager and Spectrometer
SMM Submillimetre galaxy
SNIFS SuperNovae Integral Field Spectrograph
SPIFFI Spectrometer for Infrared Faint Field Imaging
SWIFT Short Wavelength Integral Field Spectrograph
TF Tully–Fisher
TIGER Traitement Intégral des Galaxies par l’Etude de leurs Raies
TTF TAURUS Tunable Filter
UFTI UKIRT Fast-Track Imager
UIST UKIRT Imager-Spectrometer
UKIRT United Kingdom Infrared Telescope
VIMOS Visible Imaging Multi-Object Spectrograph
VIRUS Visible Integral-Field Replicable Unit Spectrograph
VLT Very Large Telescope
VPH Volume Phase Holographic
WFPC Wide-Field Planetary Camera
WHAM Wisconsin H-alpha Mapper
WHT William Herschel Telescope
WiFeS Wide-Field Spectrograph
WIYN Wisconsin–Indiana–Yale
WYFOS Wide-Field Optic Spectrograph
1. Introductory review and
technical approaches
MARTIN M. ROTH
1.1 Preface
The topic of the XVII IAC Winter School is ‘3D Spectroscopy’: a powerful astronomical
observing technique, which has been in use since the early stages of the first prototype
instruments about a quarter of a century ago. However, this technique is still not con-
sidered a standard common user tool among most present-day astronomers.
3D Spectroscopy (hereafter ‘3D’) is also called ‘integral field spectroscopy’ (IFS),
sometimes ‘two-dimensional’ or even ‘area’ spectroscopy, and commonly also ‘three-
dimensional’ spectroscopy; in other areas outside astronomy it is called ‘hyperspectral
imaging’, and so forth. It is already this diversity in the nomenclature that perhaps
reflects the level of confusion. For practical reasons, the organizers of this Winter School
and the Euro3D network (which will be introduced below) have adopted the terminology
‘3D’, which is intuitively descriptive, but, as a caveat early on, is conceptually mislead-
ing if we restrict our imagination to the popular picture of the ‘datacube’ (Figure 1.1).
Although this term will commonly be used throughout this book, we need to point out
for the reasons given later in the first chapter that the idealized picture of an orthogonal
cube with two spatial, and one wavelength, coordinate(s) is inappropriate in the most
general case.
Whatever the terminology, it is the aim of this Winter School to help alleviate the
apparent lack of insight into 3D instrumentation, its use for astronomical observations,
the complex problems of data reduction and analysis, and to spread knowledge among
a significant number of international young researchers at the beginning of their career.
The training of early stage researchers has always been the rationale in the tradition of
the well-established series of Instituto de Astrofı́sica de Canarias (IAC) Winter Schools,
but this is the first time in which a consortium of 11 leading European research institutes
appears as co-organizer, represented by the ‘Euro3D’ Research Training Network (RTN) –
which may well be recognized as the significant impact expected from this school. Euro3D
was funded by the European Commission (EC) from July 2002 to December 2005 under
1
2 Martin M. Roth
λ
Monochromatic maps
x
y
Visualization
Figure 1.2. Participants of the Euro3D Kickoff Meeting, held at IAC, Tenerife, 2–4 July 2002.
investments have been realized at almost all major observatories, with a clear technolog-
ical lead in Europe. Disappointingly, however, the expertise within the user community
for the reception of the data from such instruments was found to be very limited. It was
confined almost entirely to members of the groups that have built the 3D instruments.
Therefore, and in response to the EC requirement to demonstrate ‘the scientific, tech-
nological, or socio-economic reasons for carrying out research in the field covered by the
Figure 1.4. The three major principles of operation of present-day IFUs (Source: J. Allington-
Smith).
RECTANGLE
SQUARE
PIXSIZE2
PIXSIZE1
PIXSIZE1
HEXAGON CIRCLE
PIXSIZE1 PIXSIZE1
Figure 1.5. Different types of spaxel geometries: square, rectangular, hexagonal, bare fiber
(circular).
analysis and interpretation of their data, however, such instruments will also be men-
tioned briefly below.
Figure 1.6. Field distortion in the fore-optics of an IFU affects the spatial sampling.
100
-enhanced AR coating
90 Broadband AR coating
Front-illuminated
80
70
60
E( )
50
Measured
40
30
20
10
0
300 400 500 600 700 800 900 1000 1100
Wavelength (nm)
Figure 1.7. Fairchild CCD 4K × 4K, 15 µm pixels (left). The QE of modern CCDs can be made
quite high, very nearly approaching 100% (right).
due to the manufacturing process individual fibers cannot be arranged to arbitrary pre-
cision (for a real fiber bundle IFU, see Figure 1.23). This is another reason why the
popular description of IFS as ‘like a CCD [charge coupled device] with a spectrum in
each pixel’ is misleading. Even if the manufacture of an IFU allows a perfectly regular
sampling pattern to be created, e.g. in the case of a hexagonal lens array, the sam-
pling is not necessarily orthonormal. Moreover, real optical systems create aberrations
and, sometimes, non-negligible field distortions (Figure 1.6), which also means that the
set of spectra extracted from the detector does not sample an orthogonal FoV on the
sky. Furthermore, the sampling method may be contiguous (e.g. lens array) with a fill
factor very close to unity, or non-contiguous (e.g. fiber bundle) with a fill factor of sig-
nificantly less than 100%. In all of these cases, it is possible to reconstruct maps at a
given wavelength through a process of interpolation and, repeating this procedure over
all wavelengths, to convert the result into a datacube. Note, however, that interpolation
often produces artifacts and generally involves loss of information.
are no longer popular, except for read-noise limited applications with very low photon
count rates. An interesting new development is the photon-counting LLCCD, which uses
an on-chip avalanche gain technology for low noise at high frame rates (Jerram et al.,
2001).
In the near infrared (NIR) wavelength regime there has been enormous progress in
detector technologies, providing large area infrared (IR) arrays with pixel resolutions of
up to 2K × 2K and excellent electro-optical performance, and even larger arrays being
under development. A mosaic of 2 × 2 such buttable arrays is shown in Figure 1.8. While
a strong driver for this evolution has come from the development of the next generation
space telescope – the James Web Space Telescope (JWST) – it is also the demand for
detectors from NIR instrumentation for ground-based astronomy that has made NIR
astronomy a rapidly growing field.
The number of spectra delivered by present-day 3D spectrographs is typically in the
range of several hundreds up to a few thousand: Visible Imaging Multi-Object Spectro-
graph (VIMOS) IFU: 6400 (Le Fevre et al., 2003). The corresponding number of spaxels is
disappointingly small if one compares with direct imaging detectors: even the simple first
generation commercial CCD imagers were offering, for example, 384 × 576 ≈ 221 Kpixels:
a factor of 34 larger than the number of spaxels of the largest IFU to date. So what is
it that limits the size of an integral field spectrograph in terms of number of spatial and
spectral elements?
First of all, IFS is obviously more complex than direct imaging: for any spatial element,
a full spectrum must be accommodated. If n, m, k are the number of spatial resolution
elements in x and y, and of the spectral resolution elements, respectively, then the total
number N of required detector pixels is N = n × m × k. In comparison with direct imaging
where k = 1, IFS requires at least a factor of k more detector pixels. Clearly, instrumen-
tation for IFS is expensive in terms of detector space.
Another important factor is the spectrograph optical system. The input focal ratio as
determined by the telescope and any fore-optics, the magnification from the spectrograph
slit to the detector pixel size, as well as the total length of slit determine the spectrograph
parallel beam size and thus the overall size of the optical system.
For a large number of spectra, a wide-field optical system is typically required. Such
optical systems are challenging in terms of technical feasibility (controlling optical aber-
rations), and cost. As an example, Figure 1.9 shows the layout of the PMAS (Potsdam
Multiaperture Spectrophotometer) spectrograph optics, which is optimized to cover the
wavelength range 350−900 nm (Roth et al., 2005). It is worthwhile noting that the lens
sizes are 200 mm in diameter. While this size represents the state-of-the-art for modern
8 Martin M. Roth
Figure 1.9. PMAS fiber spectrograph as an example for an IFS optical system. Left: a cross
section, showing the layout of collimator, grating and camera. The parallel beam measures
150 mm in diameter. The magnified detail to the lower left illustrates the fiber interface. Right:
photograph of the camera, which has a total mass of 60 kg.
refractive optical systems for spectroscopy, the apochromatic correction from the ultra-
violet (UV) to the NIR is unusual and quite demanding. The cost of such a system is
high, mainly because of the use of expensive materials, e.g. CaF2 , which are required for
high throughput and the large wavelength coverage. Instruments that are not designed
for general-purpose operation but rather for a specific science case can often be built
simpler and with less effort.
Given the total number of spaxels of the IFU, the critical question of how to allocate
the corresponding spectra on the detector arises. Obviously, the total number of detector
pixels is related to the number of spatial and spectral resolution elements like:
ΣPixels ≥ NSpaxels × NSpectra (1.1)
In most conventional 3D spectrograph designs the detector space is not contiguously
filled with spectra, basically in order to avoid cross-talk between adjacent spectra and
to facilitate data reduction. Novel slicer designs like, for example, SINFONI (Eisenhauer
et al., 2003) or the OSIRIS lens array IFU (Larkin et al., 2006) have, however, demon-
strated that this is not a necessary condition. Figure 1.10 shows several basic ways of
arranging the spectra. Generally, a trade-off needs to be made between maximizing the
number of spaxels, or the number of spectral pixels.
Introductory review and technical approaches 9
Figure 1.11. Typical CCD frame of an arc lamp exposure from a fiber bundle IFU. The direc-
tion of dispersion is horizontal. The magnified insert demonstrates how the instrumental profile
is given by the FWHM (full-width-at-half-maximum) of the projected fiber size on the detector
( ≈ 4 pixels).
a) b)
Figure 1.12. Examples for instrumental efficiency as a function of wavelength for different
gratings. a) PMAS with lens array IFU, b) PMAS with PPak IFU.
Another method has emerged with the introduction of optical fibers in that the image
information at the telescope focal plane is scrambled, and thus the one-to-one corre-
spondence between points on the image plane and on the detector plane is lost. This
principle is employed in an even more rigorous way in 3D spectrographs of the TIGER-
type (TIGER = Traitement Intégral des Galaxies par l’Etude de leurs Raies; named
after the prototype; Bacon et al., 1995), where a simple fore-optics system followed by a
lens array creates a tiny image of the entrance pupil (telescope aperture) for each lens.
These so-called micropupils are designed to typically measure a few tens of micrometers
in diameter. When fed to the collimator input plane of an imaging spectrograph, the fam-
ily of micropupils practically operates like the apertures of a conventional multi-object
spectrograph (e.g. the EFOSC-type), with the one significant distinction, namely that it
is the telescope pupil rather than the telescope image that forms the multi-slit configu-
ration. The important consequence is that the light distribution of the slit illumination
no longer matters in terms of wavelength definition, which is an important argument for
precision spectroscopy. The principle is not entirely new, but was first employed using
a ‘Fabry lens’ for photoelectric photometers in order to eliminate the photometric noise
introduced by seeing (Michlovic, 1972). Another variant of the TIGER micropupil tech-
nique is employed in fiber-optical 3D spectrographs, where the fiber bundle is connected
to a lens array.
a)
b) c)
Figure 1.13. a) Raw detector frames from VIMOS IFU; b) SAURON; c) SPIFFI. See text for
explanation.
discern. Finally, the SINFONI example illustrates the appearance of slicer data, which
at first glance looks like noise. Closer inspection reveals the staggered slit arrangement
as a vertical pattern right from the center. It is obvious that this IFU type produces the
highest density of spectra on the detector, making it a good choice for IFUs requiring a
large FoV and many spatial elements.
Figure 1.14. Example: extraction of MPFS spectra from raw CCD frame.
the direction of dispersion is approximately 2 pixels, and given the more or less Gaussian
profile of the spectrograph point-spread function (PSF), there is considerable overlap
between adjacent spectra. The average inter-spectra minimum intensity is ≈ 65% of the
average peak intensity, while the spectrum-to-spectrum intensity variation is > 20%.
Ideally, the data reduction algorithm must:
(1) find the exact location of all spectra (tracing);
(2) extract the flux per wavelength bin for each spectrum;
(3) compensate for different amount of cross-talk introduced by the overlap of spectra;
(4) apply a ‘fiber-flat-field’ correction to compensate for the variation of throughput
over different fibers; and
(5) rebin all spectra to a common wavelength scale, before the set of spectra can be
used to reconstruct maps, etc.
All or at least part of these steps are actually necessary at the telescope during observing if
the observer wishes to find out whether the IFU was centered correctly on target, whether
the expected signal-to-noise (S/N) has been reached, and whether other indicators of data
quality are met. Note that the data reduction process is a non-trivial exercise, which is
one of the reasons why it took quite a long while for 3D spectroscopy to develop from
the stage of prototypes to the break-through as common user technique.
1 2 # l ID
o rd ord oup axe
r
Co Co G Sp
ID th x
u m ion X eng nde
tr t A l i
ec lec SP ec art
Sp Se N Sp St
Figure 1.15. The Euro3D data format (from Kissler-Patig et al., 2004).
this latter step of interpolation, merely assuming that the basic steps of data reduction
have been applied to remove the instrumental signature, but otherwise presenting the
data as a set of spectra with corresponding positions on the sky. Figure 1.15 illustrates
the spaxel-oriented approach of the Euro3D (E3D) FITS data format.
Data analysis
3D data analysis comprises the inspection of data quality and the determination of scien-
tifically useful physical quantities. It needs to be clearly distinguished from data reduc-
tion, which should rather be considered instrumental, and as such (almost) part of the
data acquisition process. Some conventional goals of 3D data analysis are: reconstruc-
tion of maps at certain wavelengths, line-ratio maps, velocity fields, co-adding/binning
of spectra, sky/background subtraction, spectrophotometry, etc. These topics will be
addressed extensively later in this volume. At this point it is, however, appropriate to
point out that there is a considerable level of complexity involved if one compares these
with well-established astronomical observing techniques. Although significant progress
has been made over the years, it seems fair to state that 3D spectroscopy as a general
technique is still in its infancy. In an attempt to popularize 3D spectroscopy in the astro-
nomical user community, the Euro3D consortium has made an effort to develop concepts
to facilitate the process of generating science results from 3D data. A key instrument
to this end is a 3D visualization tool, which can be described as a general graphical
user interface to 3D data. The E3D programme (Figure 1.16) as a result of this effort,
developed by S. Sánchez, is extensively described in this book.
Figure 1.16. The Euro3D Visualization Tool (Sánchez, 2004; Sánchez et al., 2004a).
a) b)
Figure 1.17. a) ‘A posteriori’ advantage of 3D; b) unlike in slit spectroscopy, it is not necessary
to center the object exactly within the aperture of the instrument.
!
!
atmospheric refraction;
!
spatial binning, low surface brightness;
!
differential spectrophotometry;
!
crowded field 3D spectroscopy;
ultra-deep faint object 3D spectroscopy.
Slit effects
Slit effects are essentially light losses caused by the finite width of a spectrograph slit.
The severeness of the effect is a function of seeing FWHM in relation to the slit width
(Jacoby and Kaler, 1993), and a function of wavelength. Bacon et al. (1995) have pointed
Introductory review and technical approaches 15
a) b)
Figure 1.18. Slit effects: a) a wide slit collects a large fraction of target flux; however, it
also collects a large amount of background light. b) A narrow slit minimizes the background
contribution (and also increases spectral resolution), but at the same time reduces the recorded
target flux. No such effect occurs in an IFU.
a) b)
Figure 1.19. a) In slit spectroscopy, the effect of atmospheric refraction can be minimized by
aligning the spectrograph slit with the atmospheric spectrum (parallactic angle). b) Because the
seeing FWHM varies with wavelength, flux calibration is affected by systematic errors; however,
this is not so for IFUs.
out that a surface brightness gradient across the slit in the direction of dispersion causes
a bias in the instrumental profile, which in turn translates into an error on the wavelength
scale. 3D spectrographs based on the micropupil principle are totally unaffected by this
problem, which is why these instruments are superior tools to measure subtle velocity
field variations on small scales, e.g. in the nuclear regions of galaxies.
Atmospheric refraction
Atmospheric refraction is due to dispersion in the atmosphere, which causes a different
angle of deflection as a function of wavelength for rays at oblique angles of incidence, i.e.
for objects at airmass > 1. For point source images, the effect is equivalent to creating
an extended spectrum instead of a well-defined spot in the focal plane of the telescope.
The spectrum is extended in the direction of the parallactic angle, i.e. a great circle
through the zenith and the target, as projected on the sky. Atmospheric refraction is
a function of air temperature, pressure and relative humidity. The effect is described
quantitatively by Filippenko (1982), who advised that a spectrograph should be placed
such that the slit orientation is along the parallactic angle or, in other words, that all
light of the atmospheric spectrum falls into the slit aperture. Since, however, seeing is
a function of wavelength, different fractions of total stellar flux are recorded along the
atmospheric spectrum. Due to this fact, spectrophotometric calibrations by means of a
flux standard are affected by systematic errors whenever the standard is observed at an
airmass that is different from the airmass of the object. None of these problems is an
issue for 3D spectroscopy, at least whenever an IFU with a fill factor close to unity is
employed. Contrary to a slit, the effect of atmospheric refraction translates merely into
16 Martin M. Roth
E 131 4612 Å
b) E 131 7440 Å
Figure 1.20. Maps of standard star exposure at two different wavelengths: a) blue; b) red. The
offset due to atmospheric refraction and the wavelength dependence of the seeing FWHM are
clearly visible. Data obtained with the GMOS-IFU at GEMINI-S (courtesy of J. Turner).
Figure 1.21. Atmospheric refraction in the datacube picture: two maps at two different wave-
lengths are shown as orthonormal grids. Owing to atmospheric refraction, the centroid of a
stellar image experiences a translation from one wavelength to another. Note that, in the most
general case, two exposures obtained at different times (and, therefore, at different airmass and
hour angles) are affected by different translation vectors. Hence, the combination of two datasets
requires a transformation to a common coordinate system, i.e. a correction of the translation
due to refraction.
Differential spectrophotometry
As outlined already above (under the headings of Slit effects and Atmospheric refrac-
tion), 3D spectroscopy is, in principle, the ideal method to perform spectrophotometry.
Spectrophotometry, however, is severely compromised under non-photometric observing
conditions. Provided the observer is in the position to find a secondary or tertiary flux
standard near the position of the target, one can apply the technique of ‘differential spec-
trophotometry’, analogous to the development of differential photometry (Barwig et al.,
1987). Using this technique, it is possible to continue observing even through thin clouds
and to compensate for transmission variations through the simultaneous observation of
the standard star. Novel 3D spectrographs with large FoV, or instruments with deploy-
able IFUs, or 3D spectrographs incorporating a direct imager for field acquisition and
guiding are ideally suited to adopt this technique.
a)
b)
Figure 1.22. Crowded field 3D spectroscopy: the LBV candidate star B416 in M33. a) Hubble
Space Telescope (HST) field, used for deconvolution technique; b) reconstructed map as inferred
from INTEGRAL data. See Fabrika et al. (2005).
for an actual second generation very large telescope (VLT) instrument by Bacon et al.
(2002). It is worth mentioning that the data reduction and stacking process of ground-
based 3D data are significantly more complex than the same process for direct images
from space. Note in particular the complication of different amounts of atmospheric
refraction for datasets obtained at different times (see Figure 1.21). However, with the
validation of concepts for faint object 3D spectroscopy with real observations and stacking
data from many hours of exposure time (e.g. Bower et al., 2004; Roth et al., 2004b), the
development of MUSE (Multi Unit Spectroscopic Explorer, a dedicated faint object 3D
spectrograph for the European Southern Observatory – VLT) is now underway (Bacon
et al., 2004). The deep surveys planned with this instrument are expected to discover
faint Lyα emitters within the 1 arcmin FoV at unprecedented flux levels down to ≈
3 × 10−19 erg s−1 cm−2 .
of a similar fiber bundle, coupled to a dedicated fiber spectrograph with a modern CCD
detector (INTEGRAL at WHT; Arribas et al., 1998a; 1998b) yielded high-quality data
of 3C 120, which allowed the application of image-processing techniques to accurately
disentangle the radio jet emission from the underlying galaxy (Sánchez et al., 2004b).
Another experiment was the DensePak-1 IFU for the KPNO 4 m telescope (Barden
and Scott, 1986), which also had the shortcomings of an early prototype. However, the
early experiments stimulated a whole series of new fiber bundle IFU developments at
different telescopes; for example, DensePak-2 (Barden and Wade, 1988) was a much
improved follow-up at the KPNO 4 m, featuring a rectangular fiber bundle with a total
of 49 fibers over a FoV of 16′′ × 19′′ and a projected fiber diameter of 2.1′′ on the sky;
SILFID (Spectrographe Integral Linarisation par Fibres de l’Image Directe) for the CFHT
(Vanderriest and Lemmonier, 1988); HEXAFLEX/WHT (Arribas et al., 1991), 2D-FIS
(2-Dimensional Fiber ISIS System) WHT (Garcı́a Marı́n et al., 1994); HEXAFLEX-II/
NOT (Garcı́a Marı́n et al., 1994); ARGUS (Vanderriest, 1995). Although fiber bundle
IFUs suffer from incomplete sampling and sometimes irregular arrangement of the fibers,
they have an intrinsically simple structure and can be fitted as add-on modules to exist-
ing bench-mounted spectrographs, which probably explains why they represent the most
common type of IFS in the early days of 3D spectroscopy. A more recent interesting appli-
cation is the SparsePak IFU at WIYN (Wisconsin–Indiana–Yale and National Optical
Astronomy Observatory; Bershady et al., 2004).
It is worth mentioning that when the topic of IFS had gained considerable momentum
in the 1990s, a dedicated International Astronomical Union (IAU) Colloquium was orga-
nized in 1994 (Compte and Marcelin, 1995), after which, superseding the initial phase
of prototypes, a first generation of common user 3D spectrographs emerged at 4 m-class
telescopes.
One of the prominent examples is INTEGRAL at the WHT (Arribas et al., 1998a;
1998b). Unlike several of the early prototypes, INTEGRAL is an optimized fiber bun-
dle IFU, attached to the WYFOS (Wide-Field Optic Spectrograph) fiber spectrograph
at WHT (Bingham et al., 1994). INTEGRAL features three rectangular fiber bundles
at different magnifications for the purpose of accommodating different observing condi-
tions in terms of seeing, and different user requirements (spatial resolution versus FoV).
Each bundle is also furnished with a set of separate fibers for sky background subtrac-
tion. The three bundles have parameters as follows (FoV/projected fiber diameter): SB1
7.8′′ × 6.4′′ /0.45′′ ; SB2 16.0′′ × 12.3′′ /0.9′′ ; SB1 33.6′′ × 29.4′′ /2.7′′ . Besides the different
Introductory review and technical approaches 21
Broken ( 78)
1.485 mm
Broken ( 167)
2.718 mm
(6.69' )
6.731 mm
Broken ( 30)
(12.25' )
(30.34' )
Broken ( 17)
600 µm
200 µm
100 µm (2.70' )
(0.90' )
1.742 mm (0.45' )
(7.95' ) 3.643 mm
(16.42' ) 7.636 mm
(34.42' )
choice of IFU magnification, INTEGRAL offers a variety of grating options with dif-
ferent dispersion, wavelength coverage and blaze thanks to the many modes available
in WYFOS (Wide-Field Optic Spectrograph). This property makes INTEGRAL a very
useful common user instrument as it allows observers to a large extent to optimize the
instrumental setup and accommodate to their specific scientific problems.
Fiber bundle IFUs have the advantage that they are conceptually simple and that they
can be attached a posteriori to existing (preferably fiber) spectrographs. It is therefore
probably no surprise that the very first IFUs were indeed based on this principle. However,
a few shortcomings soon led to alternate designs, in particular the inevitable drawback
of fiber bundles having fill factors of less than unity. The incomplete sampling does not
only influence the spatial resolution of seeing-limited observations, but also gives rise to
photometric inaccuracies when observing spectrophotometric flux standards: whenever
an appreciable fraction of standard star flux falls into the gap space between adjacent
fibers, the amount of lost light cannot be straightforwardly accounted for. A way to
overcome this shortcoming was to use an optical device that had already been proposed
by Courtès (1960), namely a lens array in the optical plane of the telescope. While
22 Martin M. Roth
the principle of operation is explained in more detail in Section 1.4.2, suffice it to say
for the purpose of this overview that a contiguous array of microlenses of square or
hexagonal shape does not suffer dead space between adjacent spatial elements, unlike
a conventional fiber bundle. Such a device was therefore thought to be ideal for the
highest spatial resolution of IFS. Using the principle of micropupil imagery and suitable
dispersion in combination with a waveband selecting filter, it could also accommodate a
large density of spectra on the detector, i.e. a larger number of spatial elements. Despite
these advantages, it was not until 1987 when the first lens array IFU ‘TIGER’ saw first
light at the CFHT (Courtès et al., 1988; Bacon et al., 1995). Other implementations
were shortly thereafter developed for the Selentchuk 6-m telescope (Afanasiev et al.,
1990) and in Kyoto (Ohtani et al., 1994). The TIGER instrument was not only ground-
breaking in terms of technological innovation, but also quite successful with scientific
achievements, e.g. the kinematic study of the double nucleus in M31 (Bacon et al., 1994),
to name just one example. A subsequent follow-up development built by the same group
in Lyon, France, was the OASIS (Optically Adaptive System for Imaging Spectroscopy)
IFS, which was optimized for very high spatial resolution in combination with the PUEO
(Probing the Universe with Enhanced Optics) adaptive optics system at CFHT. Although
adaptive optics (AO) operation was difficult to accomplish with natural guide stars only,
this development has pointed to what is now generally thought to have become a leading
track towards novel spectroscopic instrumentation at large and extremely large telescopes
(ELT), i.e. AO-assisted integral field spectroscopy (see below). It has even been claimed
that modern adaptive optics spectroscopy in general only makes sense when performed
with an IFU.
Although not strictly speaking a common user instrument, a further development of
the TIGER principle has led to what became extremely successful in terms of science
output, namely the SAURON IFS at the WHT (Bacon et al., 2001). This instrument
was very specifically designed and is operated for a particular science case, i.e. absorp-
tion or emission line kinematics and stellar population analysis in elliptical galaxies. An
overall simple opto-mechanical layout with practically no moving parts has made the
instrument very stable and reliable for what is essentially a single-purpose mission; it
has accomplished this with superb results (e.g. de Zeeuw et al., 2002). In the trend of
an increasing number of 3D spectrographs becoming available at 4 m class telescopes,
SAURON may well be counted in the category of post-prototype instrumentation, albeit
having the status of a private instrument.
In the process of the development of the baseline designs for IFS as discussed above –
i.e. fiber bundle versus lens array – it became clear that the two are largely complementary
to each other: large wavelength coverage, relatively coarse spatial sampling and moderate
number of spectra, versus moderate wavelength coverage, optimal spatial sampling and
large number of spectra. However, it was found that for certain observational problems
it would be desirable to combine the superior spatial sampling properties of lens arrays
with the large wavelength coverage of fiber bundle IFS. As a result, hybrid ‘lens array –
fiber bundle’ IFUs were developed (see Section 1.4.3). The first implementation of this
type of IFU happened in the MPFS (Multi-Pupil Field Spectrograph) instrument at the
prime focus of the Selentchuk 6 m telescope in 1997. Other IFSs using this principle are
SPIRAL/AAT (Kenworthy et al., 2001), PMAS/Calar Alto (Roth et al., 2005; Kelz et al.,
2006) and several other instruments at 8 m-class telescopes as discussed below.
Fiber and lens-array-based designs have governed the layout of the first generation
of IFSs. Given the rather limited number of spaxels in comparison with direct imaging
detectors, and also considering the inefficient use of detector space, the slicer principle
was introduced with the ‘3D’ instrument (Weitzel et al., 1996). This IFS was designed
Introductory review and technical approaches 23
Figure 1.27. PMAS at the Calar Alto 3.5 m telescope features two IFU options: a fibercoupled
lens array IFU, and a bare fiber bundle IFU.
24 Martin M. Roth
a) b)
Figure 1.28. a) Compact GMOS-IFU. b) VIMOS-IFU, the largest IFU to date, providing a
total of 6400 spectra.
for the J, H and K bands in the near infrared (NIR). Since the available detector size
was small (256 × 256 pixels), it was felt that the number of fiber or lens array spectra
that could be projected onto the available chip would be too small to warrant the effort
of building a conventional IFU. As an alternative, the slicer method generates a stack of
long-slit spectra, corresponding to a series of parallel slices in the image plane. Contrary
to the fiber and lens array IFUs, the sky is thus imaged directly onto the detector, whereas
in the former case the fiber end faces or micropupils are projected onto the focal plane.
The detector area is therefore optimally filled with contiguous spectra and no space is
lost due to gaps. The optomechanical design of slicers is less critical at long wavelengths,
but more difficult in the optical or UV as far as performance in terms of image quality,
edge effects and stray light is concerned. A few more NIR slicer instruments were built
for 4 m-class telescopes, but none for operation in the optical (PIFS/Hale 5 m, Murphy
et al., 1999; UIST/UKIRT, Ramsay Howatt et al., 2004).
The history of IFS as discussed so far included the very first prototypes and a first gen-
eration of facility instruments for 4 m class telescopes. The latter already represented a
significant step towards establishing 3D spectroscopy as a common user technique. In the
late 1990s the light collecting power, superb image quality and demand for new instru-
mentation of new 8–10 m class telescopes prompted yet another generation of IFSs, some
of which were designed as IFUs in the original sense, i.e. add-on modules in front of con-
ventional slit spectrographs (e.g. GMOS-IFU: Gemini Multi-Object Spectrograph – IFU;
Allington-Smith et al., 2002; VIMOS-IFU/VLT: Visible Imaging Multi-Object Spectro-
graph – IFU/VLT; Le Fevre et al., 2003), while others are dedicated 3D spectrographs
(SINFONI/VLT, Eisenhauer et al., 2003; OSIRIS/KECK, Larkin et al., 2006). These
instruments usually come with their own data reduction pipeline, which is an important
prerequisite for opening this non-trivial observing technique to the common user.
Future facilities are presently planned for ground-based and for space telescopes. The
most prominent new developments for the ground are MUSE as a second-generation
VLT instrument, and VIRUS (Visible Integral-Field Replicable Unit Spectrograph) for
HET (Hobby–Eberly Telescope). These new facilities are planned to accommodate a
total of as many as 90,000 and 30,000 spectra, respectively. As a consequence, they are
quite large and highly complex, requiring novel approaches for manufacture and integra-
tion, e.g. replicable devices which can be built in small industrial series, thus reducing
risks and saving costs. In space, JWST is planned to host – for the first time – IFUs
above the atmosphere, namely as part of the NIRSpec (Near-Infrared Spectrograph)
and Mid-Infrared Instrument (MIRI), covering the 0.6−5 µm and 25−27 µm wavelength
ranges respectively (Arribas et al., 2005). Of those future instruments, MUSE will come
Introductory review and technical approaches 25
Figure 1.29. MUSE at the Nasmyth platform of the VLT (CAD layout).
reasonably close to the ideal of a conventional imager, however creating a full one-octave
spectrum in each pixel. Over the wavelength range of 465−930 nm, MUSE will accommo-
date a square FoV of 1 arcmin with a spatial sampling of 0.2 × 0.2 arcsec2 in its wide-field
mode. The narrow-field mode will cover 7.5 × 7.5 arcsec2 with a sampling of 0.025 arcsec2 .
The instrument will be assisted by a ground-layer correcting AO system with 4 laser
guider stars and an additional natural guide star (Bacon et al., 2004).
! optical fibers;
that are in use for 3D instrumentation; namely:
! lens arrays;
! lens arrays coupled to fiber bundle;
! slicers;
! special techniques;
! non-IFS 3D instruments;
! 3D detectors; and
! figures-of-merit.
! the cladding as a low refractive index skin around the core; and
! a protective buffer normally made of some sort of plastic material, e.g. polyamide.
26 Martin M. Roth
They act as wave guides owing to the effect of total internal reflection, meaning that rays
at grazing incidence onto the core–buffer interface will be reflected back into the core.
Multiple reflections inside the core lead to a scrambling of the input light pattern. Rays
incident onto the fiber input with inclination angles larger than that given by a certain
critical angle (numerical aperture, NA) will no longer undergo total internal reflection,
but will rather be absorbed in the buffer. Figure 1.31 shows a microscopic view of the
polished front face of a bundle formed from several such fibers which were bonded together
to make up a hexagonal densest package. In this picture, the bright circular areas are the
fiber cores, surrounded by a dark ring (the cladding), and another material (the buffer).
Note that this fiber bundle IFU has an incomplete fill factor, as only a fraction of the
area is acting as a light-collecting aperture. One of the most important disadvantages
of fibers is their property of not normally retaining the input f-number of the incoming
beam (Figure 1.32). The emerging light cone has a larger opening angle than the input
beam. The effect is called focal ratio degradation (FRD). It is very undesirable as it leads
Introductory review and technical approaches 27
Figure 1.33. Several views of the fiber-optical IFU ‘PPak’ of PMAS (Kelz et al., 2006).
to overfilling the spectrograph collimator and light losses. Moreover, since the amount
of FRD depends on external parameters like mechanical or thermally induced stress, the
output f-number is not constant with time unless the fiber duct from the IFU to the
spectrograph is extremely stable. In certain cases, variable FRD is capable of creating
noise, which limits the ultimate S/N obtainable, even under conditions with a lot of
light and good photon statistics (Parry and Carrasco, 1990; Baudrand and Walker, 2001;
Schmoll et al., 2003).
In summary, fiber-optical IFUs have the following properties: Fibers allow for a very
flexible installation of an IFU, and they accommodate bench-mounted spectrographs,
in particular, which is a strong argument for stability. There are a few problematic
fiber properties, which – with proper care and a suitable design – can be handled fairly
well. Fiber-optical IFUs have typically relatively few spaxels, but – depending on the
spectrograph used – a large number of spectral bins and a wide wavelength coverage.
Fiber bundle IFU can be built with large spaxels to cover a sizeable area on the sky, which
is, for example, extremely beneficial for low surface brightness spectroscopy. One of the
more serious drawbacks for certain applications, especially when high spatial resolution is
intended, is the non-contiguous sampling due to inevitable gaps in the bundle. However,
contrary to common belief, fiber-optical IFUs can be built with high efficiency.
Magnifying lens
Macrolens array
Figure 1.36. Optical layout of a lens array with fore-optics (adapted from Dean, 2002).
area is negligible for most lens arrays. Lens arrays are therefore superior devices for high
spatial resolution studies of objects on small scales. Another significant advantage of lens
array IFUs is often overlooked, namely the principle of micropupil imagery. Instead of
imaging the dispersed spaxels of the IFU, it is the microscopic entrance pupil behind each
lenslet (= spaxel) of a lens array that is imaged onto the detector (hence the term ‘micro-
pupil’). The practical advantages of micropupils for precision spectroscopy are discussed
in Section 1.2.1.
The sketch in Figure 1.35 illustrates the optical principle, which is well known from
the technology of one-channel photoelectric photometry, where a so-called ‘Fabrylens’
images the entrance pupil of the telescope onto the photocathode of a multiplier. Instead
of having to accept the noise caused by seeing-induced image motion of a stellar image
on the photocathode, the static pupil image is unaffected by atmospheric effects or guid-
ing errors. Translating this macroscopic arrangement to the scale of a lens array in
Introductory review and technical approaches 29
combination with some fore-optics having the proper focal length, the result is a set
of microscopic pupils behind the lens array, with one micropupil for each lenslet. The
size of the micropupils is designed to match the optimal slit width of an imaging spec-
trograph with a sufficiently large FoV to encompass the physical size of the lens array.
The arrangement is then very similar to a multi-object slit spectrograph of the EFOSC
(ESO Faint Object Spectrograph and Camera) type (Buzzoni et al., 1984), except that
no punch plate is necessary as the slits are defined by the micropupils, and the micro-
pupils are aligned on a totally regular grid. In order to avoid the overlap of spectra com-
ing from neighboring micropupils, the spectral coverage per spectrum must be truncated
to approximately the pitch of the lens array, which can be done by using a narrowband
filter in the fore-optics train. Moreover, in order to make more efficient use of detector
area and to maximize the spectral coverage, the dispersive element (grating, grism) is
slightly rotated, resulting in a diagonal arrangement of spectra in the spectrograph focal
plane. Figure 1.37 shows a real example from the SAURON instrument (Bacon et al.,
2001) as magnified views from:
a) the undispersed micropupils;
b) a continuum flat-field exposure;
c) a spectral line lamp calibration exposure; and
d) an astronomical object exposure.
In summary, lens array IFUs can, in principle, accommodate a large number of spax-
els at the expense of wavelength coverage, and some possible problems with truncated
spectra near the edge of the detector. The fact that there is contiguous sampling and no
gaps makes them very suitable for high spatial resolution observations. Thanks to the
micropupil principle, the telescope focal plane is decoupled from the detector, which is
a significant advantage for precision spectroscopy. The compact geometry of the spectra
allows for compact optomechanical designs and also for highly efficient optical systems.
A certain drawback of lens array IFUs is the presence of diffraction causing crosstalk
between neighboring spectra, which occurs at different wavelengths (Lee et al., 2001;
30 Martin M. Roth
Roth et al., 2005). This undesirable effect can, however, largely be corrected for by soft-
ware in the process of data reduction.
T elescope
focal plane
Output to
spectro meter
1.4.4 Slicers
The IFU types discussed in Sections 1.4.1 to 1.4.3 have one feature in common: they
do not use detector space very efficiently. This obviously has to do with the fact that
re-imaging of fiber end faces, or micropupils, respectively, which incurs a finite image
quality, and, in fact, a broadened image of those ‘slit elements’, makes it necessary to
provide for a minimal spacing of spectra, unless the crosstalk from the broadened profiles
is not regarded as critical. If this is indeed the case, one spectrum would typically cover
four rows of spectra in the spatial direction, assuming that Nyquist sampling was given
by design. Such a configuration consumes a factor of two more CCD pixels than required
in the optimal case, where one spaxel is critically sampled, i.e. it uses just two rows of
CCD pixels. If maximization of the number of spaxels is a primary design goal, then the
slicer concept is significantly more efficient than the competing IFU types. Slicer IFUs
are conceptually based on the optical elements designed to circumvent slit losses in con-
ventional spectrographs; e.g. the Bowen–Walraven image slicer. The image slicer for 3D
spectroscopy was first introduced for the MPE (Max Planck Institut für extraterrestrische
Physik) instrument ‘3D’ (Weitzel et al., 1996), which was designed as a prototype and a
travelling instrument; it was adapted and operated at a variety of 4 m-class telescopes.
3D operates in the NIR, which precludes the use of fibers because of stability issues in a
cryogenic environment. Conversely, the slicer as a relatively small optical element is well
adapted for operation in a cryostat at low temperatures and allows for an overall compact
optomechanical design. The principle of operation of a slicer is shown in Figure 1.39. Light
coming from the telescope is deflected by a stack of tilted mirrors (the ‘slicer’ proper),
which are very narrow in one direction and quite extended in another direction. These
thin mirrors ultimately act as slits, or, in other words, as spaxels, dissecting the focal
plane into narrow and long strips. Another set of mirrors accepts the deflected beams
corresponding to each slice and projects the outgoing beams such that an (almost) linear
arrangement of the slices is created at the entrance of the spectrograph. The benefit of
such a configuration is that the telescope focal plane is directly imaged onto the detector
so that, with the choice of a proper scale, one detector pixel can be interpreted as one
spaxel on the sky. Of all IFU types, this scheme yields the highest packing density on the
detector. However, unlike with micropupil or fiber-based 3D spectrographs, slicer IFUs
do not provide image scrambling and therefore do not share the benefits as discussed
in Section 1.2.1. The first slicers were used for instruments operating in the NIR, e.g.
32 Martin M. Roth
Reflective coating
r
B
A α
α
π phase
change
α
d
Fabry–Perot
A very prominent technique, starting from the early days of IFS developments, is the
Fabry–Perot interferometer (FP). The basic principle of interference in a FP is shown in
the sketch of Figure 1.41: a cavity enclosed by a pair of high-quality plane-parallel glass
plates with a reflective coating of reflectance r is acting as a resonator, where rays of
Introductory review and technical approaches 33
Photomultiplier
tube
Mode selector
wheel
Flip in mirror
1/8 Beam
Interference filter F1
Flip in mirror
Detector
Collimator 1/7 B7 393 mm focal length Fabry–Perot Camera lens – 1/2 102 mm focal length
equal angles of incidence experience the same phase superposition, and – depending on
the optical path length difference – constructive or destructive interference. The optical
path length between adjacent rays is given by: 2AB − CD = 2d cos α. The condition for
constructive interference is then:
2d cos α = mλ (1.2)
For every wavelength λ satisfying this condition, the interferometer is transmissive, and
other wavelengths are extinct. Because of the periodicity of the cos-function there are
very many orders m with constructive interference when the spacing d is on the order
of mm. By varying d, one can tune the interferometer to different wavelengths, and
scanning over many values of d is effectively creating a spectrum. Note that in order to
avoid ambiguity owing to the many orders m, it is necessary to use a narrowband filter
to select the wavelength of interest. The width of the passband depends on the etalon
reflectance r, and it can be shown that the spectral resolving power is given by:
√
λ r
= mπ (1.3)
∆λ (1 − r)
A prominent example for a scanning FP imager is the TAURUS instrument (Atherton
et al., 1981; 1982), the layout of which is shown in Figure 1.42. Some of the instrumental
characteristics of TAURUS were as follows. Mode of operation: seeing-limited direct
imaging, FoV = 9 arcmin, spectra resolution: 200 < R < 100,000, IPCS photon counting
detector, wavelength range: 400−700 nm. General features of a scanning FP are: a large
field-of-view, high spatial resolution, non-simultaneous measurements, small wavelength
coverage (generally one spectral line), high spectral resolution.
Beam splitter
Path B
Collimator
Moveable
mirror
Path A
Camera
Detector
Magnetic field
Substrate
AIOx barrier
less straightforward than other 3D spectrographs, but it has a certain number of advan-
tages. These are quoted with several instrumental parameters of the CFHT FTS as an
example (Bohlender, 1994): the CFHT FTS is a scanning NIR instrument (1−5 µm),
having a high throughput of up to 25%, independent of wavelength. High S/N observa-
tions are supported based on a dual channel beam switching mode, which allows for an
accurate internal sky subtraction before any signal is detected. The instrumental profile
is extremely well controlled (and measurable). The wavelength calibration accuracy is
≈ 100 m s−1 , and the spectral resolution up to 500 m s−1 .
1.4.6 3D detectors
The dream of observational astronomers would be a photon counting detector that sam-
ples all parameters of an incident photon simultaneously; i.e. wavelength, phase and
Introductory review and technical approaches 35
amplitude, polarization and direction. 3D spectrographs as described above are a hum-
ble yet very powerful approximation towards this goal. However, a compact detector
having these properties would seem to be more appealing intrinsically than the rather
complex layout of the various IFU types. Such detectors do exist, and a prominent pro-
totype development is the Superconduction Tunnel Junction (STJ). This device consists
of a pair of films that are separated by an insulating layer. The films are made of super-
conducting materials like niobium, tantalum, hafnium, etc. The detector is operated at a
temperature below Tcrit ≈ 1K. Owing to the Josephson effect, there is a current through
the junction in the presence of a bias voltage and a magnetic field. The current is sup-
pressed by incident light, where an intrinsic energy resolution is provided due to the
release of a charge from the incident photon. The typical resolving power that can be
achieved, however, is R = 10−100, i.e. very low. This limitation is given by the band-gap
values attained by selected materials, for which research and development efforts are
attempting to provide further improvement. For fundamental reasons, however, spectro-
scopically interesting resolutions one order of magnitude larger than the present values are
not going to become real. Therefore, the STJ will remain a niche application that cannot
replace IFUs. Nevertheless, 3D detectors with spatial resolution have been demonstrated
as arrays of 10 × 10 Josephson cells, forming a real 3D detector. First astronomical obser-
vation with the STJ were presented by Perryman et al. (1999). Their measurements of the
Crab pulsar validated that the STJ has advantages for high time resolution multiband
imaging, especially in the UV. Other developments include so-called Kinetic Inductance
Detectors (KID), which are being developed as large devices with 100 × 100 elements and
R = 30, or Transition Edge Detectors (TES) with 8 × 8, R = 20.
In conclusion, intrinsic energy resolution detectors are still at an early stage of devel-
opment. Their cryogenics to operate at superconduction temperatures around 1 K are
complex and expensive. The devices can only be made to achieve low spectral resolution.
Although they are sensitive in the blue, they are less suitable for the visual and useless
in the NIR. A positive feature is their high time resolution.
s''
S'
Ω' Ω'' Ω∗
Ω
AS Entrance
A
slit
IFUs shows that the uncertain location of point sources over the slit perpendicular to
the slit length obviously introduces large flux uncertainties, making this technique less
competitive compared to, for example, the MUSE instrument.
Nod-shuffle 3D spectroscopy
The technique of ‘nod-shuffle’ spectroscopy is not necessarily restricted to 3D; however, it
is a very powerful option for background-limited observations of extended objects, where
no patch of empty sky can be found within the FoV of the IFU. As described by Cuillandre
et al. (1994), consecutive beam switching between object and sky pointings while shuffling
charge on the CCD detector between an object and a sky integration zone, and final
read-out after the total exposure time has been completed, is capable of eliminating
systematic sky subtraction errors in the sense of chopping technique, which are common
place in many areas of measuring techniques. A comprehensive discussion is presented
by Glazebrook and Bland-Hawthorn (2001), and experience from an implementation in
the PMAS instrument was reported by Roth et al. (2004a).
1.4.8 Figure-of-merit
In order to be in a position to objectively compare competing instrument designs, it is
useful to introduce a quantitative measure of performance. The classical measure to this
end is ‘etendue’, which is defined as:
Etendue = area × solid angle × system throughput (1.4)
Introductory review and technical approaches 37
Another measure, which takes out the system throughput and thus gauges the perfor-
mance of an IFU irrespective of the attached spectrograph is ‘grasp’:
grasp = area × solid angle (1.5)
Examples for grasp of a number of 3D instruments are presented by Bershady et al.
(2004). Another figure-of-merit, called ‘specific information density’ was proposed by
Allington-Smith (2007).
While these numbers are useful for planning the optimal design of an IFU for a given
science case, observers may also want to consider other parameters, which may be even
more relevant than a single figure-of-merit; for example, spatial resolution, spectral res-
olution, wavelength regime and wavelength coverage.
1.5 Acknowledgments
The Euro3D Research Training Network was funded by the European Commission in
Framework Programme 5 under contract number CT-2002-00305. The initiative of the
IAC, La Laguna, Tenerife, to organize the XVIIth IAC Winter School on the topic of ‘3D
Spectroscopy’ as a task of the Euro3D Work Plan, to provide funding, and to arrange a
marvelous school at the beautiful location of Puerto de la Cruz is gratefully acknowledged.
REFERENCES
Afanasiev, V.L., Vlasiuk, V.V., Dodonov, S.N., Silchenko, O.K. (1990), Preprint No.54,
Special Astrophysical Observatory, Selentchuk
Allington-Smith, J.R. (2007), in Science Perspectives for 3D Spectroscopy (ESO Astrophysics
Symposia), eds. Kissler-Patig, M., Wals, J.R., and Roth, M.M. Springer
Allington-Smith, J., Murray, G., Content, R. ET AL. (2002), PASP, 114, 892
Arribas, S., Mediavilla, E., Rasilla, J.L. (1991), ApJ, 369, 260
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B. ET AL. (1995), A&AS, 136, 189
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B., del Burgo, C., Fuensalida, J.J.
(1998a), A&AS, 136, 189
Arribas, S., Cavaller, L., Garcı́a-Lorenzo, B., Garcı́a-Marı́n, A., Herreros, J.M.,
Mediavilla, E., Pi, M., del Burgo, C., Fuentes, J., Rasilla, J.L., Sosa, N. (1998b),
in Fiber Optics in Astronomy III, eds. S. Arribas, E. Mediavilla, F. Watson, ASPCS, 152,
p. 149
Arribas, S. ET AL. (2005), Bulletin of the American Astronomical Society, 37, 1352
Atherton, P.D., Reay, N.K., Ring, J.R., Hicks, T.R. (1981), Opt En, 20 (6), 806
Atherton, P.D., Taylor, K., Pike, C.D., Harmer, C.F.W., Parker, R.N. (1982), MNRAS,
201, 661
Bacon, R., Emsellem, E., Monnet, G., Nieto, J.L. (1994), A&A, 281, 691
Bacon, R., Adam, G., Baranne, A., Courtès, G., Dubet, D., Dubois, J.P., Emsellem,
E., Ferruit, P., Georgelin, Y., Monnet, G., Pecontal, E., Rousset, A., Say, F.
(1995), A&AS, 113, 347
Bacon, R., Copin, Y., Monnet, G. ET AL. (2001), MNRAS, 326, 23
Bacon, R., Adam, G., Cabrit, S. ET AL. (2002), in Scientific Drivers for ESO Future
VLT/VLTI Instrumentation, eds. J. Bergeron, G. Monnet, Springer, Berlin, p. 108
Bacon, R., Bauer, S.-M., Bower, R. ET AL. (2004), SPIE, 5492, 1145
Balick, B., Preston, H.L., Icke, V. (1987), AJ, 94, 1641
Barden, S., Scott, K. (1986), BAAS, 18, 951
Barden, S.C., Wade, R.A. (1988), ASPCS 3: Fiber Optics in Astronomy, 113
Barwig, H., Schoembs, R., Buckenmayer, C. (1987), A&A, 175, 327
Baudrand, J., Walker, G.A.H. (2001), PASP, 113, 851
Becker, T., Fabrika, S., Roth, M.M. (2004), AN, 325, 155
Bershady, M.A., Andersen, D.R., Harker, J., Ramsey, L.W., Verheijen, M.A.W.
(2004), PASP, 116, 565
38 Martin M. Roth
Bingham, R.G., Gellatly, D.W., Jenkins, C. ET AL. (1994), SPIE, 2198, p. 56
Bland–Hawthorn, J., Taylor, K. (1994), AAO Newsletter, 71, 8
Bohlender (1994), CFHT FTS Manual
Bower, R.G., Morris, S.L., Bacon, R. ET AL. (2004), MNRAS, 351, 63
Buzzoni, B., Delabre, B., Dekker, H. ET AL. (1984), Msngr, 38, 9
Cappellari, M., Copin, Y. (2003), MNRAS, 342, 345
Compte, G., Marcelin, M. (1995), Tridimensional Optical Methods in Astrophysics, N◦ 149.
IAU Colloquium, 1994
Courtès, G. (1960), AnAp, 23, 96
Courtès, G., Georgelin, Y., Bacon, R., Monnet, G., Boulesteix, J. (1988), in Instru-
mentation for Ground-Based Optical Astronomy, Proc. 6. Santa Cruz Summer Workshop,
ed. L.B. Robinson, Springer, New York, p. 266
Cuillandre, J.C., Fort, B., Picat, J.P. ET AL. (1994), A&A, 281, 603
Dean, A. (2002), Thesis, IoA University of Cambridge
de Zeeuw, P.T., Bureau, M., Emsellem, E., Bacon, R., Carollo, C.M., Copin,
Y., Davies, R.L., Kuntschner, H., Miller, B.W., Monnet, G., Peletier, R.F.,
Verolme, E.K. (2002), MNRAS, 329, 513
Eisenhauer, F., Abuter, R., Bickert, K. ET AL. (2003), SPIE, 4841, 1548
Emsellem, E., Bacon, R., Monnet, G., Poulain, P. ET AL. (1996), A&A, 312, 777
Fabrika, S., Sholukhova, O., Becker, T., Afanasiev, V., Roth, M., Sánchez, S.F.
(2005), A&A, 437, 217
Filippenko, A.V. (1982), PASP, 94, 715
Fruchter, A.S., Hook, R.N. (2002), PASP, 114, 144
Gerhard, O., Arnaboldi, M., Freeman, K. ET AL. (2005), ApJ, 621, L93
Glazebrook, K., Bland-Hawthorn, J. (2001), PASP, 113, 197
Hill, J.M. (1988), ASPCS 3: Fiber Optics in Astronomy, 3, 77
Hill, J.M., Angel, J.R.P., Scott, J.S., Lindley, D., Hintzen, P. (1980), ApJ, 242, L69
Hook, R.N., Lucy, L.B. (1993), in Proc. Science with the HST, p. 245
Jacoby, G.H., Kaler, J.B. (1993), ApJ, 417, 209
Jerram, P., Pool, P.J., Bell, R. ET AL. (2001), SPIE, 4306, 178
Kelz, A., Verheijan, M.A.W., Roth, M.M. ET AL. (2006), PASP, 118, 129
Kenworthy, M.A., Parry, I.R., Taylor, K. (2001), PASP, 113, 215
Kissler-Patig, M., Copin, Y., Ferruit, P., Pecontal-Rousset, A., Roth, M.M. (2004),
AN, 325, 159
Larkin, J., Barczys, M., Krabbe, A. ET AL. (2006), New Astronomy Review, 50, 362
Lee, D., Haynes, R., Ren, D., Allington-Smith, J. (2001), PASP, 113, 1406
Le Fevre, O., Saisse, M., Mancini, D. ET AL. (2003), SPIE, 4841, p. 1670
Mediavilla, E, Arribas, S. (1993), Natur, 365, 420
Meisenheimer, K., Hippelein, H. (1992), A&A, 264s, 455
Michlovic, J. (1972), App Opt, 11, 490
Murphy, T.W., Jr., Matthews, K., Soifer, B.T. (1999), PASP, 111, 1176
Ohtani, H., Sasaki, M., Aoki, K., Takano, E., Kiyohara, M. (1994), SPIE, 2198, 229
Parry, I.R., Carrasco, E. (1990), SPIE, 1235, 702
Parry, I.R., Bunker, A., Dean, A. ET AL. (2004), SPIE, 5492, 1135
Perryman, M.A.C., Favata, F., Peacock, A. ET AL. (1999), A&A, 346, L30
Ramsay Howatt, S.K., Todd, S., Leggett, S. ET AL. (2004), SPIE, 5492, p. 1160
Roth, M.M., Becker, T., Schmoll, J. (2000), ASPCS 195: Imaging the Universe in Three
Dimensions, 122
Roth, M.M., Fechner, T., Wolter, D., Kelz, A., Becker, T. (2004a), in Scientific Detec-
tors for Astronomy, The Beginning of a New Era, eds. Amico, P.; Beletic, J.W.; Beletic,
J.E., 371
Roth, M.M., Becker, T., Kelz, A., Schmoll, J. (2004b), ApJ, 603, 531
Roth, M.M., Kelz, A., Fechner, T., Hahn, T., Bauer, S., Becker, T., Böhm, P., Chris-
tensen, L., Dionies, F., Paschke, J., Popow, E., Wolter, D. (2005), PASP, 117, 620
Sánchez, S.F. (2004), AN, 325 (2), 167
Sánchez, S.F., Becker, T., Kelz, A. (2004a), AN, 325 (2), 171
Sánchez, S.F., Garcı́a-Lorenzo, B., Mediavilla, E. ET AL. (2004b), ApJ, 615, 156
Schmoll, J., Roth, M.M., Laux, U. (2003), PASP, 115, 854
Si’lchenko, O., Afanasiev, V. (2000), A&A, 364, 479
Introductory review and technical approaches 39
Tully, R.B. (1974a), ApJS, 27, 415
Tully, R.B. (1974b), ApJS, 27, 437
Tully, R.B. (1974c), ApJS, 27, 449
Vanderriest, C. (1980), PASP, 92, 858
Vanderriest, C. (1995), in Tridimensional Optical Spectroscopic Methods in Astrophysics, IAU
Coll. 149, ASPCS 71, p. 209
Vanderriest, C., Lemonnier, J.-P. (1988), in Instrumentation for Ground-Based Optical
Astronomy, Proc. 6. Santa Cruz Summer Workshop, ed. L.B. Robinson, Springer, New
York, 304
Walsh, J.R., Roth, M.M. (2002), Msngr, 109, 54
Weitzel, L., Krabbe, A., Kroker, H., Thatte, N., Tacconi-Garman, L.E., Cameron,
M., Genzel, R. (1996), A&AS, 119, 531
Wells, D.C., Greisen, E.W., Harten, R.H. (1981), A&AS, 44, 363
Wilkinson, A. Sharpless, R.M., Fosbury, A.E.M., Wallace, P.T. (1986), MNRAS, 218,
297
Williams, R.E., Blacker, B., Dickinson, M. ET AL. (1996), AJ, 112, 1335
2. Observational procedures
and data reduction
JAMES E.H. TURNER
2.1 Introduction
In this chapter, I give an introduction to observing with integral field units and performing
basic reduction of the resulting data, prior to scientific analysis. After briefly considering
the context of the lectures, I begin by discussing strategies for observing. This is followed
by a short tutorial on sampling theory and its application to integral field unit (IFU) data,
before continuing with an overview of the requirements for each stage of data reduction.
I finish by considering the data reduction process as a whole, along with associated issues
such as error propagation and file formats.
2.2 Background
Techniques for integral field spectroscopy (IFS) have been in development for at least
two decades (Vanderriest, 1980).1 During the 1980s–1990s, numerous prototype IFUs
and even a few public instruments were deployed at observatories and used for scientific
work. Nevertheless, IFS has only become widely available at major telescopes during the
past five years or so, following two centuries of slit spectroscopy. Experience in observing
with IFUs and processing the data is just starting to become commonplace within the
community, but will be spread more widely by the current generation of postdoctoral
and student astronomers.
In terms of data reduction and analysis, IFS poses some non-trivial new requirements.
The most obvious factor is the introduction of 3D datasets to mainstream optical and
near-infrared (NIR) (as opposed to radio) astronomy. Although older scanning methods
such as Fabry–Perot interferometry produce higher-dimensional datasets, these tech-
niques are relatively specialized by comparison. IFS is an extension of traditional spec-
troscopy, covering a range of spatial and spectral sampling regimes, increasingly so as
larger-format units become available. This creates a need for non-specialists to apply
spectroscopic and image processing to 3D datasets. Perhaps a more important issue than
dimensionality, however, is that IFUs scramble spatial information in the process of map-
ping three dimensions onto a 2D detector. To add to this, the spatial pixels of an IFU
rarely form a square grid to begin with. Special software is therefore needed to recover
spatial or 3D images. Depending on the processing strategy, this may affect multiple data
reduction stages. Finally, a third consideration in dealing with IFU data is homogene-
ity, given variations in transmission, spectral resolution or background counts across the
field. This leads to additional flat-fielding requirements, for example, compared with a
simple aperture.
Although instruments with IFUs have become more common in the past few years,
the availability of matching general-purpose software has lagged behind somewhat. A
key task of the Euro3D Research Training Network (2002–05) has been to create data
visualization and analysis tools to help address this issue, as well as a standard file format
for working with reduced IFU data. Other recent progress includes adaptation of image-
viewing programs to work with 3D data. The proliferation of high-level environments such
1
Ignoring the even earlier use of image slicers to avoid slit losses (Bowen, 1938).
40
Observational procedures and data reduction 41
Figure 2.1. Sketch of the relative atmospheric transmission as a function of wavelength in the
optical and near infrared. Some common photometric bandpasses are labelled with the appropri-
ate letters. (Based on a figure from a US Naval publication, reproduced at http://wikipedia.org.)
as Python and interactive data language (IDL), with multi-dimensional array handling
capabilities, also provides some flexibility for manipulating 3D datasets. There is still
progress to be made, however, before the same range of tools exist for working with IFU
data as for slit spectroscopy.
Figure 2.2. The image slicing IFU of the GNIRS spectrograph at Gemini South (right),
mounted alongside various slit masks in a motorized sliding mechanism that inserts the appro-
priate mask into the beam.
More often than not, this translates into small field sizes (arcseconds) compared with
typical long slits or imagers (arcminutes or even degrees). Some IFUs, on the other hand,
are designed with coarse spatial pixels and/or short spectra, giving a larger field (∼tens
of arcseconds) and potentially much better sensitivity for low surface brightness objects,
instead of resolving fine spatial detail (e.g. Bershady et al., 2004). Sometimes it is possible
to choose between multiple spatial scales.
A second common characteristic of IFUs is that they introduce non-uniformities over
the field, as mentioned in Section 2.2, especially with optical fibres. Together with a small
field size or coarse sampling this makes dithering and mosaicking important strategies
for observing with IFUs, both to average over variations and to cover a larger area
or improve the sampling (see Section 2.3.2 Dithering and mosaicking, and Section 2.4).
Moreover, the IFU field response has to be ‘flat fielded’, in addition to that of the detector,
unlike the step response of a slit.
Finally, it is worth noting that high-spatial-resolution IFUs are a natural companion for
adaptive optics systems, both scientifically (many spectroscopic projects using adaptive
optics benefit from 2D spatial coverage) and technically (the difficulty of using a slit at
such small spatial tolerances). This is another source of special requirements for observing
and data reduction, albeit outside the scope of the current chapter.
Nod to sk Nod to sk
ffsets on source
bserve bserve
blank sk blank sk
xN xN
ffsets on source
xN 1-4
bserve standard
at position B
Overview
Figure 2.3 is a sketch of a typical on-sky observing procedure at optical wavelengths. After
centring the target on the IFU, some number of exposures of its spectrum are taken, often
interspersed with flat and/or arc lamp calibration spectra (where flexure is important). If
the target is extended enough to cover the whole field, it may be necessary to observe it
alternately with blank sky; however, this is usually avoidable with separate fibres or lenses
at some distance from the field centre, sampling the background contemporaneously
with the object. Smaller spatial offsets between target exposures are sometimes used for
mosaicking or averaging over flat-field variations.
Figure 2.4 is a sketch of a typical on-sky observing procedure in the near-infrared.
Unlike in the optical case, it is usual to alternate exposures on the science target with
blank sky spectra (perhaps in an ABBA sequence, to reduce offsetting overheads) in order
to sample the object and background with the same optical path and detector pixels.
Given the abundance of telluric emission lines in the near infrared, there is typically less
need to take arc lamp spectra for wavelength calibration, but frequent flat fields may
still be required in the presence of instrumental flexure. Usually a separate standard
star observation is taken – close to the science target in both airmass and time – to
calibrate the telluric absorption features found at longer (red to infrared) wavelengths.
44 James E.H. Turner
Being compact, the star can normally be offset between two or more positions within the
IFU field for sky subtraction, without spending time off source.
Whatever the observing wavelength, after some number of hours (depending on flexure
or telescope guiding), a re-acquisition step may be necessary in order to maintain good
centring on the target. This is more important if the telescope guider does not compensate
for changing atmospheric dispersion between the observing and guiding wavelengths.
Nevertheless, centring is likely to be less critical than for a typical narrow slit.
More often than not, some additional calibration data are required that need not be
taken on the sky with the science target (Table 2.1). This may include flat-field or arc-
lamp frames if flexure does not make the illumination dependent on time or telescope
pointing. For optical CCDs, bias exposures are needed, to find the zero point of the data.
For infrared arrays or older CCDs, dark exposures may be used to remove counts due
to thermally excited electrons. Twilight sky flats provide the spatially flattest source of
strong illumination if artificial sources are not adequate alone. If absolute flux calibration
is not needed, a spectrophotometric flux standard star can be observed within a reason-
able time of the science data (days or even weeks) to calibrate the sensitivity spectrum of
the telescope+instrument. (In the near infrared, the telluric standard may also serve this
purpose.) Depending on the project, additional stellar observations may be needed as
spectral templates for velocity or line index measurements. Either kind of standards may
be taken in poor observing conditions if absolute fluxes are not required. Finally, special
calibrations are required for some instruments, for example to calibrate the alignment of
the IFU with the detector.
Acquisition
Blind telescope pointing is rarely accurate enough to centre a target perfectly within a
small IFU field, so acquisition often involves iterative centring at the required position –
usually the middle of the IFU – as for a slit. At each step, an exposure is taken, the
target position is measured and the telescope (or in some cases the IFU) is offset slightly
to adjust the pointing.
Measurement of the target position can be done either with a normal imaging camera
that is fixed with respect to the IFU (perhaps a different mode of the same instrument)
or by reconstructing a 2D image of what the IFU itself is looking at (Figure 2.5). In the
former case, the target is offset until it reaches pixel coordinates known to correspond
to the IFU centre, based on prior measurements. This requires that the position of the
IFU on the camera is repeatable, without too much flexure or other sources of significant
error. In the latter case, the best sensitivity is achieved by taking images undispersed
through the IFU (or, if applicable, using the wavelength of an emission line). The 1D slit
or micropupil spots are then rearranged to form a 2D image. If saturation is a concern
Observational procedures and data reduction 45
Figure 2.5. a) Acquiring a target object into a box marking the IFU position in the GNIRS
direct
imaging field; b) section of an image of the same target, taken undispersed through the IFU;
c) a reconstruction of the IFU field, made by stacking the image slices from the exposure in the
previous panel.
or the spectrograph does not allow taking undispersed images, spectra can be summed
over wavelength before reconstructing a 2D image, or a 3D datacube can be constructed
first and then collapsed in wavelength (if not too slow). Finally, if flexure is of concern
and initial pointing is poorer than a few arcseconds – or the target is faint or difficult
to identify – the two approaches can be combined by using a direct camera image to
place the target within the IFU field first and then fine-tuning the pointing by taking
exposures through the IFU.
When combining exposures from an instrument with large spatial pixels it may help to
have repeat acquisitions (e.g. from different nights) accurately aligned; if spatial informa-
tion is missing due to the coarse pixel size, co-alignment of data during later processing
is likely to be problematic (Section 2.4). Differences in atmospheric dispersion between
exposures, however, tend to limit how well alignment can be maintained over a range of
wavelengths, so some error is often unavoidable. For spatially well-sampled IFU data, on
the other hand, one can correct for both pointing offsets and relative changes in atmo-
spheric dispersion afterwards; thus, in many cases periodic re-acquisition is less important
than for a slit, where miscentring on the aperture is difficult to correct afterwards and
can lead to loss of signal.
Figure 2.6. Spectra from separate, interleaved blocks of object and sky fibres in the GMOS
IFU (this detector subsection overlaps 2 of a total of 30 blocks of fibres).
At optical wavelengths, the sky can usually be sampled at the same time as the object,
using one of several methods. If the target is compact enough, one can use blank sky from
the edges of the IFU or nod up and down the field in order to subtract pairs of frames.
With multiple targets one might dither around the field and later co-add exposures
with rejection to remove the objects. Often, however, the instrument is designed with a
separate background field or sky fibres to take care of the problem. These are placed far
enough away from the main science field (e.g. a few arcminutes away) to fall on blank
sky in most cases, or at least to contain negligible signal from the outer regions of an
extended target, compared with the science field. If the target is particularly extended,
however (or maybe the sky field is contaminated by an object), it is necessary to nod
off the science target completely from time to time and expose on blank sky elsewhere.
Figure 2.6 shows an example of separate object and sky spectra from a fibre IFU.
In the near infrared, standard practice is to nod to sky, usually every other exposure, to
allow removal of telluric and thermal emission and detector dark current by subtracting
pairs of raw exposures. Since sky lines are stronger than in the optical and vary on
timescales as short as a few minutes, it is usual to nod to sky more frequently for optimal
subtraction. For compact targets such as standard stars, we can again nod within the
IFU field to subtract pairs of exposures without spending time off source, but more often
than not half of the observing time is spent on blank sky.
At non-thermal wavelengths (∼1 µm) and high enough spectral resolution, it may be
sufficient for some projects to spend all of the observing time on source and later mask
out or interpolate over sky lines, after subtracting separate dark frames instead of sky.
Integration times
IFS exposure times are mainly determined by the same factors as for other spectroscopic
modes. For faint targets, the minimum exposure is, as always, limited by the need to inte-
grate long enough for the background noise to overcome detector read-out noise. With
IFUs it often takes longer to reach the same number of counts per pixel as for a slit.
One reason is that high spatial or spectral resolution IFUs have smaller apertures than a
typical slit, being able to sample finer detail without losing light overall (but, of course,
spreading the light over more apertures ultimately leads to higher read noise). IFUs also
place extra optics in the beam, with a corresponding loss of efficiency (although through-
puts of up to ∼ 90% have been achieved, compared with a slit in the same spectrograph).
A third factor is that IFU optics sometimes introduce extra magnification, which again
means spreading the light over more pixels. On the other hand, IFUs with large spa-
tial elements – designed for high sensitivity rather than fine sampling – can reach the
Observational procedures and data reduction 47
read-noise limit very much more quickly. For very bright objects, the minimum exposure
time is sometimes set by instrumental factors such as detector read-out speed.
Maximum exposure times are affected by several factors. In the infrared we have to
start a new exposure often enough to allow for variations in sky lines (on multiple charac-
teristic timescales) for accurate subtraction. For bright targets, one can take short expo-
sures of a few seconds to a few minutes, to sample the fast telluric variations. For fainter
targets, longer exposures of 10–30 minutes are used, to average over short-timescale vari-
ations without going so long as to be affected by slower changes. Exposures much longer
than these usually lead to poor sky subtraction, although the stability varies from night
to night. When working between sky lines at higher resolutions, the subtraction accuracy
may be less important. Another common time limitation for bright objects (or perhaps
bright sky lines) is the need to avoid reaching non-linear count levels or saturating the
detector capacity altogether with too many photo-electrons. Finally, there are various
other practical reasons for dividing up observations, such as allowing for changes in flex-
ure between the slit and detector or the telescope image and IFU, the use of repeated
samples to help remove cosmic rays, spatial dithering on the source or simply avoiding
excessive time loss in the event that something goes wrong with an integration.
Given the above considerations, typical spectroscopic exposure lengths for scientific
targets range from a few minutes up to 60 minutes (in the optical) or ∼ 20 minutes (in
the near infrared).
2
Sub-pixel dithering is referred to as ‘subsampling’ in the original lectures, but that was
perhaps not the best choice of terminology, given prior use of ‘supersampling’ for the same idea.
48 James E.H. Turner
aligning exposures. If larger mosaics are needed, one either has to rely on the hardware for
accurate movements or use multiple reference peaks with well-known relative positions
to calculate the offsets.
For IFUs with coarse spatial pixels, sub-pixel dithering may help avoid loss of spatial
information and the resulting processing artifacts. This idea is best known in connec-
tion with the Drizzle technique for HST Wide-Field Planetary Camera (WFPC) data
(Fruchter and Hook, 2002). A likely example would be stepping by one half or one
third of a fibre, lenslet or image slice between exposures (much smaller increments do
not necessarily gain extra detail, given their size relative to the smoothing effect of the
larger pixel area, but may improve stability when combining data with offset errors; see
e.g. Lauer, 1999). It is important to know the offsets accurately in order to interleave
exposures correctly. Unlike the Hubble Space Telescope (HST), ground-based observa-
tories suffer from variable seeing and cloud, which are likely to limit how accurately
a higher-resolution image can be reconstructed from multiple poorly sampled datasets.
To the author’s knowledge, sub-pixel dithering has not yet been well tested with IFUs,
although a 3D-capable version of the Drizzle algorithm has recently been written by
National Optical Astronomy Observatory (NOAO) and is being tested for use with the
Near Infrared Integral Field Spectrograph Integral Field Unit (NIFS IFU) on Gemini
North.
At near-infrared wavelengths, dithering and mosaicking could potentially allow spend-
ing a larger fraction of observing time on source than for a single target pointing plus
sky. The most common observing sequence for nodding to sky is probably sky–object–
object–sky, repeated as many times as necessary. This optimizes the signal-to-noise ratio
for faint targets, since there is equal background noise from sky and object exposures.
When offsetting, however, data are combined with relative shifts, so a single sky exposure
can be subtracted from multiple object exposures and the results co-added without
using the same measurement repeatedly at any given position. Thus a sequence such as
sky–object–object (repeated) can be used. The signal-to-noise ratio is the same as when
using equal numbers of object and sky exposures, but only if data do not need binning
or fitting spatially. This does not, of course, remove the need to observe sky frequently
enough to sample temporal variations. In practice, the most conservative schemes (i.e.
sky–object– or sky–object–object–sky– repeated) tend to give the best results.
Flat-fielding
As for other observing modes, flat-field frames are needed as a reference for correcting
instrumental efficiency variations (with x, y or λ). Since an IFU introduces structure in
the focal plane, there are two components to the overall flat-field response: the usual
detector flat, including pixel-to-pixel variations and other multiplicative features, and an
IFU flat, accounting for differences in transmission between the optical fibres, lenslets
or image slices that make up the IFU, or along the length of slices. Variations over
the IFU can be due to several factors, such as fibre stresses (focal ratio degradation),
alignment variations between optical elements, optical bonding, differences in surface
roughness and diffraction losses. The IFU and detector are often flat-field corrected as
a single system, but can be treated separately if necessary due to flexure. In order to
have the correct spatial and spectral characteristics simultaneously, a final flat can be
constructed from more than one source (see below). The spatial structure of the flat cor-
responds largely to the IFU, whilst spectral variations are mainly due to the detector and
spectrograph.
Spectrally, (detector) flats require a dispersed source of illumination with a smooth
continuum: usually a bright calibration lamp. Dispersing the light mimics science data,
Observational procedures and data reduction 49
a) b)
c) d)
Figure 2.7. Some example IFU flats: a) a calibration lamp spectrum from the GNIRS image
slicer; b) a twilight sky spectrum, showing telluric structure; c) small subsection of a lamp
spectrum from the GMOS fibre IFU, covering two blocks of fibres with a gap between them;
and d) a vertical cross-section of the previous panel, showing the throughput differences and
overlaps between optical fibres.
allowing for any wavelength dependence of the flat-field pattern (e.g. due to quantum
efficiency or fringing). An intrinsically smooth spectrum is needed to distinguish small-
scale flat-field structure from the shape of the spectrum itself, which is typically fitted
and removed during processing. The detector flat may incorporate a fringe pattern due
to internal optical interference, especially for CCDs at longer wavelengths. For the IFU
flat component, calibration lamp data can also be used to measure differences in spectral
response (rather than just overall throughput) between IFU elements, if such differences
happen to be significant. If separate detector and IFU flats are needed, on the other
hand, one could take lamp exposures through a clear long-slit aperture, matching the
spectral setup of the science data without the IFU in the focal plane.
Spatially, a flat illumination source is needed for correcting IFU structure and any
vignetting by the spectrograph. The most uniform reference is the twilight sky (although
the sky has a usable brightness level for an interval of only a few minutes, twice per day;
for high spatial or spectral resolution IFUs, the brightness may even be optimal before
sunset or after sunrise, making for long observing nights). Lighting a blank, neutral
area inside the telescope dome with appropriate lamps is another common technique for
achieving relatively flat illumination. Over the small area of a typical IFU field, however,
even an inbuilt calibration lamp may be sufficiently uniform for the spatial content of
the flat, if designed appropriately. Figure 2.7 shows some examples of different types of
flat-field spectra.
50 James E.H. Turner
Given that individual IFU spectra are much narrower than for a long-slit, flexure
perpendicular to the dispersion axis is more of an issue; a shift of only a few pixels
could correspond to the width of a fibre/lenslet or a large fraction of an image slice.
It is therefore important that flats are well matched to their corresponding scientific
exposures. Often flexure between the IFU and detector is significant, in which case the
safest strategy is to take frequent flats (e.g. once per hour) between science data, or
at least to take them at the same telescope pointing. As well as matching the detector
illumination accurately, this provides a clear reference for tracing IFU elements in the
science spectra. In the absence of flexure, flats can be taken before or after night-time
observations and applied to data from throughout the night. If some optical element
(such as the tilt of the disperser) moves non repeatably, flats or arcs should be taken
before changing the instrument configuration.
As long as they match the flexure of science data, lamp flats taken through the IFU can
be used to correct both IFU and detector variations together. Where the lamp exposures
are not spatially uniform enough on their own, a separate twilight observation can be
used to apply a flatness correction. For IFU flats that do not match science data, a
detector-only flat should also be obtained, allowing them to be reduced and shifted into
alignment or extracted. Likewise, image slicer twilights may need matching lamp flats to
remove flexure-dependent detector variations that differ from science data.
Wavelength calibration
To determine the wavelength accurately at each detector pixel, a spectrum containing
well-known features at zero redshift is needed, covering the range of the detector. The
features must both be recognizable (by their relative or absolute positions and strengths)
and have accurately catalogued wavelengths. We can then measure their pixel coordinates
and interpolate to find wavelengths at intermediate positions. For IFUs, each separate
1D fibre/lenslet or point in a 2D slitlet spectrum needs calibrating.
Often the best reference source is an arc lamp. Some common examples are argon,
copper-argon, thorium-argon, xenon and krypton lamps, which have many strong lines
with precise laboratory measurements (e.g. Figure 2.8). In the red or near infrared, an
alternative is to use telluric emission lines in the raw science spectra, with the advantage
that flexure is then not an issue. Where this is not an option (such as for high resolutions
or short exposures at thermal wavelengths), well-defined telluric absorption bands can
sometimes be used instead.
Typically, at least one arc exposure is taken in order to derive a detailed wavelength
solution, including non-linear terms. Sometimes this is observed during the day time or
twilight. Where there is flexure, any small zero-point shifts in wavelength can often be
corrected using even a single strong telluric line in the raw science spectra. Sometimes,
however, there are no suitable sky lines at all, especially for high dispersions in the
blue. If wavelength calibration is important in those cases, any flexure must be tracked
using frequent arc spectra between science exposures (ensuring that arcs are taken before
changing the disperser tilt, if it is non-repeatable). Occasionally, multiple arc lamps may
be needed to obtain enough lines over certain short wavebands.
Telluric calibration
At far red to infrared wavelengths, scientific data tend to be affected by telluric absorption
lines. The common photometric bandpasses (e.g. I, z, J, H, K, L, M) are defined in
spectral regions with relatively good atmospheric transmission, but still contain many
minor absorption features due to water vapour and other molecular species. Moreover,
Observational procedures and data reduction 51
a)
b)
Figure 2.8. a) A subsection of a typical copper-argon arc spectrum, taken with the GMOS
optical fibre IFU; b) an argon arc from the GNIRS near-infrared image slicer.
there is sometimes occasion to observe in between these clean regions, perhaps to measure
a strong emission line that is redshifted from the visible.
For many purposes, telluric absorption left in scientific spectra would be problematic.
Telluric lines can be confused with real stellar features, for example, when using an
automatic algorithm to measure velocities. More importantly, telluric lines can overlap
real spectral features, changing their profiles so that spurious line widths, centres or
strengths would be measured from the raw data.
In order to correct for telluric absorption, a star of known spectral type is observed with
little or no intrinsic absorption at wavelengths of scientific interest; typically, this is an
A-type star (Figure 2.9), which has an intrinsic spectrum that is almost featureless apart
from hydrogen lines. The amount of absorption in telluric bands scales with airmass and
varies in time, particularly with changes in water vapour. Telluric standards are therefore
observed immediately before or after the corresponding science object, with coordinates
chosen to match the airmass, or zenith distance, at which it is observed. Where the range
of airmass variation during a science observation is small (e.g. < 0.3 airmasses), a single
star can be observed, matching the average airmass. For longer observations, the range
of airmass can be bracketed with telluric standard spectra taken before and after.
If the spectral profile of the instrument varies substantially across the IFU field, one
might take telluric spectra with the standard star at multiple positions on the IFU to
attempt to calibrate different regions separately. For image slicers, point-like and diffuse
52 James E.H. Turner
HD 22789
Figure 2.9. K-band spectrum of an A-type star with telluric absorption lines. With the notable
exception of the Brackett-γ line at 2.17 µm, most of the structure in the spectrum is introduced
by the atmosphere. Normalization is arbitrary.
image features will produce slightly different spectral profiles anyway, having different
shapes across the width of an image slice; this is likely to be a relatively minor source of
error, however, especially after summing over the telluric observation spatially.
Flux calibration
To allow meaningful comparison of fluxes between different wavelengths – or with pre-
vious datasets – a stellar calibration spectrum is normally needed. Often this is just a
relative calibration of the instrumental efficiency variation with wavelength, but where
necessary it may also serve as an absolute throughput or sensitivity measurement for the
instrument+telescope+sky.
Relative flux calibration is the spectral equivalent of flat-fielding on large spatial scales
(whilst flat fielding itself takes care of smaller spectral structure from pixel-to-pixel differ-
ences and fringing). It is needed for taking the ratios of lines at different wavelengths, or
measuring the true continuum slope of a target, for example. Given photometric observ-
ing conditions, flux calibration additionally allows the brightness of a source or spectral
feature to be measured in absolute units (such as W m−2 Å−1 arcsec−2 ). Alternatively,
line strengths can be measured relative to the continuum, as equivalent widths, avoiding
dependence on accurate flux calibration.
To derive the sensitivity spectrum for a set of scientific observations, the reference star
needs to be observed with the same instrument configuration and have a known intrinsic
brightness as a function of wavelength. At visible wavelengths, numerous spectrophoto-
metric standard stars are available, with wavelength-tabulated fluxes in the literature
(e.g. Oke, 1990). In the near infrared, one often just uses a convenient star of well-known
spectral type and with a reliable broad-band magnitude measurement; the intrinsic con-
tinuum can then be modelled with a black-body curve for the appropriate temperature
and magnitude. Figure 2.10 is an example of a near-infrared absolute throughput spec-
trum, derived from a G-type star.
Observational procedures and data reduction 53
50
40
H-band throughput (%)
30
20
10
0
15,000 15,500 16,000 16,500 17,000 17,500
Wavelength (Angstroms)
Figure 2.10. Absolute throughput spectrum for the NIFS integral-field spectrograph at an H-
band setting; the curve was derived from a G-star observation by comparing the observed counts
with a black-body model of the star’s intrinsic flux.
Figure 2.11. Sketch illustrating the problem of colour-dependent slit losses due to atmospheric
dispersion (differential refraction). Although a stellar image may be well centred on the slit at
the central wavelength (middle circle), the same is not true for the extremal wavelengths unless
the slit is rotated to the parallactic angle, along the direction of atmospheric refraction. This
effect scales with airmass and is strongest at blue wavelengths, in the optical.
Flux standards can be taken one or more times during a given observing run (a short
enough period that the instrument and telescope characteristics will not change sig-
nificantly). They need not be observed at the same time as scientific spectra, unless
absolute spectrophotometry is critical or the same data are to be used for both telluric
and flux correction. Where absolute flux measurements are not required, standards can
be observed through cloud, which to first order is neutral in colour.
Because an IFU normally captures all the light from an incident stellar image, without
slit losses, observing standards is relatively straightforward. Absolute fluxes can be mea-
sured without any special setup and, in principle, a single observation may be used for
both telluric and flux calibration. For long-slit spectroscopy, on the other hand, absolute
flux calibration requires using a special wide slit to capture all the light, resulting in a
lower spectral resolution than for scientific data; since a telluric spectrum must match
the scientific resolution, this has to be taken separately. Moreover, atmospheric disper-
sion causes colour-dependent throughput losses for narrow slits, unless observing at the
parallactic angle (Figure 2.11); thus even relative measurement of a sensitivity spectrum
is easier and more accurate with an IFU.
54 James E.H. Turner
Detector bias
Raw measurements from a CCD detector include an electronic bias level. Since this
varies slightly over the array, a separate zero point needs to be determined for each
pixel, to allow calculation of the true counts accumulated above the bias. This simply
involves taking a few very short exposures (ideally of zero duration, i.e. immediate read
out), in the dark. The resulting images directly provide the zero-point value, with no
photoelectrons stored, for each pixel. These exposures include read-out noise, of course,
so several are taken and later averaged together, to reduce the noise to a level that will
not degrade scientific data when subtracted. A typical bias frame looks flat, with random
variations of just a few counts superimposed; depending on how well the electronics are
isolated from electrical noise sources, the image may also include some faint structure.
The bias level produced by detector electronics can drift with time, but is usually
stable enough for reference exposures to be taken only occasionally, during the daytime.
Sometimes the detector controller also creates an overscan region for each exposure,
by continuing to read out the detector after shuffling and retrieving all the accumulated
charge. This adds extra blank image columns, for monitoring variations in the overall bias
and read-out noise. If necessary, the overscan can be used to make an overall zero-point
correction, on top of the pixel-to-pixel differences from bias exposures.
Infrared arrays are normally read out by measuring the difference in counts between
the start and end of each exposure. This is possible because each pixel can be read out
rapidly without affecting the stored charge, whereas CCDs are read out destructively, by
shuffling the charge from one column to the next and finally off the detector. Thus, the
bias level is removed automatically for infrared observations and does not need measuring
separately.
Dark current
Over the duration of a long enough exposure, detector pixels accumulate some electrons
due to thermal excitation and array defects, as well as from incident photons. Although
detectors are cooled with nitrogen or helium to minimize this thermal current, cooling
is limited by the need to avoid poor quantum efficiency at very low temperatures. For
modern CCDs, the effect may nevertheless be negligibly low (e.g. 1 e− hour−1 pixel−1 ).
For infrared arrays, on the other hand, dark current is more important and tends to affect
some pixels much more than others. This creates variations in the raw data that obscure
real features (Figure 2.12). A reference is therefore needed for distinguishing dark counts
from those due to photons from a target object.
For any given exposure time that is used in observing, the dark current component
of images can be estimated by exposing the detector in complete darkness for the same
length of time. Multiple exposures are usually taken, as for bias frames, to account for
read noise and to allow removal of cosmic rays. Since, however, the exposures may be
much longer than biases (and may therefore be best taken at night or in twilight if light
leaks are an issue), it is not always practical to observe more than a few frames.
For scientific observations with separate, equal sky exposures, the same dark current
is present in both object and sky data. Thus in the infrared, subtracting object and sky
pairs removes dark current automatically from those spectra, without the need for special
calibrations. This usually does a good job of cleaning up most hot pixels, leaving just a
statistical increase in noise; any pixels that remain affected after subtraction are masked
out during reduction. Separate dark calibrations are required, however, in cases where
background subtraction is attempted without nodding to sky, or where sky frames will
be scaled to match on-target data more accurately.
Observational procedures and data reduction 55
Figure 2.12. Section of a raw scientific exposure from an HgCdTe array. Most of the numerous
bright specks superimposed on the target and sky spectra are ‘hot’ pixels, with high dark
current.
For infrared calibration data, matching darks are usually required unless the exposures
are short enough not to have significant dark counts. For flat fields of more than a
few seconds, it is common to take pseudo-darks in the same configuration as the flats
themselves, but with the lamp turned off. For arc spectra, darks are needed if uncorrected
hot pixels could otherwise be confused with emission lines in a 1D section.
In summary, darks are often needed at infrared wavelengths for calibrations, but are
optional (and time consuming) for scientific data, depending on the reduction method.
With modern CCDs, darks are often not needed at all (but since separate sky frames are
usually not available as a reference, they would be more critical if required).
(where the multiplicative |a| just preserves the unit area of δ when scaling the coordinate
argument by the same factor). Thus, sampling f (x) at regular intervals of a is equivalent
to multiplication with a comb function:
"x# ∞
$
III f (x) = |a| f (an) δ(x − an). (2.4)
a n = −∞
Observational procedures and data reduction 57
f(x)
III (x/a)
III(x/a) f(x)
x
a
Figure 2.14. Sampling a continuous function, f (x), at regular intervals of a, by multiplication
with a comb function.
Clearly, III f is a series of impulses, normalized to the sampled values: it contains the
same information from f (x) as would point-like measurements (Figure 2.14).
The comb function is a useful tool for representing sampling because its Fourier trans-
form is also a comb. Working with frequencies in reciprocal units of x (rather than 2π/x)
gives:
"x# F
III ↽ a III(au).
⇀ (2.5)
a
According to the Convolution Theorem, multiplication of functions in the spatial
domain is equivalent to convolution of their Fourier transforms (and vice versa). Thus
the representation of sampled data in frequency space is:
" "x# #
F III f (x) = a III(au) ∗ F (u). (2.6)
a
This is the frequency spectrum of the original, continuous f (x), replicated infinitely at
intervals of 1/a (Figure 2.15); convolving F (u) with each δ(u − u′ ) simply shifts the origin
to u′ .
A function is said to be ‘band limited’ if its Fourier transform is zero above some cut-
off frequency, uc . In Figure 2.15 and Equation 2.6, the replicas of F (u) do not overlap if
f (x) contains a maximum frequency component of uc ≤ 1/2a; thus we can separate out
F (u) from the transform of the sampling pattern and recover the original f (x) without
any loss of information. If the replicas do overlap, on the other hand, the function cannot
be determined uniquely.
58 James E.H. Turner
u
1 uc
a
Figure 2.15. The Fourier spectrum of a uniformly sampled function, III(x/a)f (x), is the Fourier
spectrum of f (x), repeated at regular intervals of 1/a.
III(au) * F(u)
u
Figure 2.16. Given the Fourier transform, III(au) ∗ F (u), of a function sampled above the
Nyquist rate, the transform, F (u), of the original, continuous function can be recovered by
multiplication with a box function, Π(au).
Having reconstructed the original analogue signal, we can then resample it arbitrarily
onto a convenient grid (such as a regular datacube, for IFU data).
3
The box cut-off can be anywhere between uc and 1/2a (if different), but the latter is
convenient because it is known and fixed for a particular sampling.
Observational procedures and data reduction 59
Figure 2.17. The sinc function, sin(πx)/πx, which peaks at unit value at x = 0 and oscillates
towards small values at infinity, with zeros at integer x.
Figure 2.18. Reconstruction of a continuous function from regularly spaced samples, by convo-
lution with a sinc function. Each grey curve is a sinc function centred at the appropriate sample
point and normalized to the corresponding sample value (cf. Figure 2.14). The thicker black
curve shows the overall sum as a function of x.
u
1 uc
a
Figure 2.19. When f (x) is undersampled, such that it contains frequency components > 1/2a
for a sampling interval of a, replicas of F (u) overlap in the Fourier spectrum of the sampled
data.
Equation 2.8 specifies f (x) at all intermediate points between the samples. This is
possible because the number of frequency components was finite to begin with, so f (x)
can be represented by a corresponding number of basis functions. The basis could instead
be a Fourier series, but finding the coefficients would then be more complex than above:
sinc functions are a natural basis, because the required coefficients are simply the sample
values.
2.4.2 Undersampling
For a given sample separation, a, if f (x) contains frequencies greater than the Nyquist
frequency, 1/2a, it is said to be undersampled. In the Fourier spectrum of the sampled sig-
nal, F (u) then overlaps its neighbouring replicas (Figure 2.19); the spectrum is effectively
60 James E.H. Turner
fn(x)
x
a
Figure 2.20. An undersampled frequency component and its lower frequency alias have the
same values at regular sampling intervals, making them indistinguishable. The Nyquist, or crit-
ical sampling, criterion requires the sampling rate to be high enough with respect to the signal
bandwidth that such ambiguities cannot occur.
wrapped around or reflected at 1/2a, such that power at frequencies above this limit is
aliased to lower frequencies. The aliasing can be thought of as beating between high-
frequency components of the signal and the sampling period itself. On a regular grid,
high- and low-frequency pairs can appear identical at every sample point, even though
their values in between are quite different (Figure 2.20). The only way to distinguish
between these aliases is to disallow one of them (the higher frequency in our case); hence
the requirement for band limitation.
Where undersampling occurs, the transfer of power from higher to lower frequen-
cies creates artifacts in the reconstruction of f (x) from samples. Any super-Nyquist
frequencies are not only unmeasurable, but contaminate otherwise good data. We can
end up with less useful information than if f (x) had been smoothed to a lower resolu-
tion/bandwidth before sampling.
With irregular sampling patterns, the similarity between alias frequencies is broken – as
can be seen from Figure 2.20 – so there is no direct transfer of power from one frequency
to another. Nevertheless, information is still lost if the sampling rate is too low, which
unavoidably leads to reconstruction error. Moreover, if the number of available samples is
too small, the problem of finding a set of basis coefficients may become ill conditioned. In
general, the theory of reconstruction from irregular samples is considerably more involved
than that discussed here (see references in Section 2.4.4, Irregular sampling).
Further information on regular sampling, including proofs of the theorems relied on
here, can be found in Bracewell (2000, or earlier editions).
Figure 2.21. Diffraction of a plane wavefront by a finite aperture. The resulting interference
produces a far-field intensity distribution similar in cross-sectional form to the sketch on the
right-hand side (where intensity increases to the right).
( 2 Jπr(πr))
2
1
1.0
0.5
r
1.22
1.02
Figure 2.22. Cross-section of the Airy intensity pattern produced by diffraction of a plane wave
by a circular aperture.
where D is the mirror diameter, λ is the wavelength of light, θ is the angular displacement
in radians from the central direction of propagation and J1 is a first-order Bessel function
of the first kind. A derivation can be found in Goodman (1968) or Hecht (1987).
The finite image of a point source produced by a given optical system is known as its
point-spread function (PSF). For a diffraction-limited telescope, the PSF is therefore the
Airy pattern described above (ignoring a smaller correction for the central obscuration).
For a seeing-limited telescope, the PSF is a superposition of atmospheric speckle images
that tends towards a Gaussian-like form over long enough (∼ 60 s) exposures.4 Since the
PSF describes the distribution of light from any point-like element, the 2D image formed
by a telescope is equivalent to an infinitely detailed projection of the sky, convolved with
the PSF.
The Fourier power spectrum of the PSF is known as the modulation transfer func-
tion (MTF). This describes the sensitivity of the system to different spatial frequencies,
4
In practice, this form is often better described by a Moffat profile (Moffat, 1969), which is
centrally similar to a Gaussian curve, but allows for broader wings.
62 James E.H. Turner
a) b)
Figure 2.23. The MTF of a telescope is described by the self-convolution of the transmission
over its aperture (so there is always some high-spatial-frequency cut-off).
convolution with the PSF being equivalent to multiplication by the MTF in the fre-
quency domain. Since a telescope’s diffraction PSF is the squared Fourier transform
of its aperture transmission, the Convolution Theorem also tells us that the MTF is
given by the self-convolution of the aperture transmission (Figure 2.23). It follows that,
for a finite aperture, telescope images must be band-limited, with some high-frequency
cut-off.
For a circular mirror of diameter D, the cut-off frequency turns out to be D/λ cycles
per radian (see Goodman, 1968). To satisfy the Sampling Theorem, the detector pixel
spacing must therefore be less than λ/2D radians. The FWHM of the Airy pattern is
1.02 λ/D radians, so in terms of the PSF the critical pixel spacing is 0.49 × FWHM.
Thus we confirm the two-pixel (really 2.04 pixel) sampling criterion. Strictly speaking,
however, this applies only to diffraction-limited imaging.
For seeing-limited images, we can approximate at least the bright core of the PSF by a
Gaussian curve with a FWHM typically an order of magnitude greater than the diffrac-
tion limit. The Fourier transform of a Gaussian curve is also Gaussian, so is infinitely
extended. Thus seeing-limited images do not have a strict frequency cut-off, other than
that imposed at much smaller spatial scales by the diffraction limit. Theoretically, we
would therefore need a FWHM of something like 20 times the pixel size to reach the
Nyquist rate. Fortunately, in practice, we can approximate critical sampling with a less
stringent criterion.
The Gaussian distribution and its Fourier transform can be written
2 1 −π (u /a) 2
F
e−π (ax) ⇀
↽ e . (2.10)
a
√ '
For unit a, these have a standard deviation of 1/ 2π and a FWHM of 2 (loge 2)/π ≈
0.939. Although both curves are infinitely extended, their values die away quickly after
a few standard deviations. If a Gaussian PSF is imaged with a FWHM of n pixels, then,
rewriting its transform in terms of the FWHM using na = 0.939, we have
2
u2
F (u) = 1.06ne−3.56n , (2.11)
with u in cycles per pixel and a standard deviation of σ = 0.375/n. For a FWHM of
n = 2 pixels, the Nyquist frequency of u = 0.5 cycles/pixel therefore corresponds to 2.7σ,
where the Gaussian MTF is at only 2.8% of its peak value. Super-Nyquist frequencies are
therefore heavily suppressed in amplitude, to < 3%. The total integrated power that is
aliased back into the main spectrum is only 0.7%. Thus, even for seeing-limited images,
the ‘2 pixel per FWHM’ criterion is good enough in most cases. A pixel density slightly
greater than this will reduce aliasing to very low levels.
Observational procedures and data reduction 63
2.4.4 Practical issues
Finite detector
The above discussion outlines how the Sampling Theorem relates to imaging, but several
idealizations still need justification. Perhaps most obvious is the fact that sampling is
done with ∼ 103 detector pixels or ∼ 101 spatial IFU pixels per dimension, rather than
an infinite comb function.
A finite detector image is equivalent to a comb function multiplied by the telescope
image (as previously) and truncated with a box function of width na, where n is the num-
ber of pixels and a their separation. The Fourier transform is therefore an infinite comb
convolved with F (u) (compare Figure 2.15), convolved again with a narrow sinc function
whose first zeros are at ±1/(na). This zero location corresponds to twice the Nyquist
frequency divided by n (i.e. two pixels of the discrete Fourier transform). The extra
convolution reflects the fact that a finite band-limited image is described by a finite
number of frequency components; otherwise, the transform of the sampled data would
be infinitely detailed. Thus, to first order the finite detector size makes no difference; the
transform of sampled data is still, approximately, F (u) repeated at intervals of 1/a. Con-
volving F (u) with something much narrower than itself does not extend the frequency
band appreciably, so the same sampling criterion applies.
To second order, things are a bit more complicated. The sinc function is infinite and
slowly decaying at lower power levels, so the transform of the sampled data actually has
no hard cut-off frequency. In fact, a finite function, such as a sampled image, can never
be strictly band-limited. This frequency extension is manifested as reconstruction error
towards the edges of the data, similar to the oscillation seen at the ends of a Fourier
series when approximating a sharp edge. The behavior can be improved somewhat by
suitable filtering.
Noise
Our discussion of the Sampling Theorem assumed noiseless data, so here we must consider
the effect of random errors. Poisson noise is incoherent and spread over all frequencies,
rather than contributing power at any particular frequency. Noise aliasing is not a con-
cern, because it would just fold noise back into the main spectrum, affecting the statistical
realization of the errors without systematic effect. The question is therefore how the noise
level propagates through the resampling process. When interpolating images, the value
at a particular x is a weighted sum of sinc functions, one per sample (see Equation 2.8
and Figure 2.18). Since noise terms add in quadrature, each of the original sample values
contributes to the total variance at x according to the square of its weight. Since the
values of sinc2 (x) at unit intervals add up to one (Bracewell, 2000), the noise level in
the interpolated image is the same as the original. Noise therefore gets carried through
64 James E.H. Turner
a) b)
x u
−a a 1 uc
2 2 a
Figure 2.24. a) Nearest-neighbour interpolation uses the closest sample value to each x and is
therefore equivalent to convolving a series of point samples with Π(x/a). b) The corresponding
envelope in the spectral domain is ∝ sinc(au), whose zero values coincide with the sampling
replicas of the original image spectrum, F (u).
the system without systematic effects, to first order. In practice, of course, the noise
level will vary over the image and the sinc function may allow bright peaks to dominate
further out.
Practical interpolants
Although a sinc function is the ‘ideal interpolant’, it is infinite and strongly oscillating, so
is not used in practice without modification. Discontinuities such as edges and bad pixels
can cause severe diffraction-like ‘ringing’ effects (Gibbs’ Phenomenon). Tapering the sinc
function with an envelope such as a Gaussian or cosine multiplier provides a more robust
and computationally feasible solution. This type of interpolant is good for well-sampled
data, albeit relatively slow to compute unless look-up tables are used. It still tends to be
more sensitive to spikes and discontinuities than other interpolants, but the envelope can
be tuned to make a trade-off between suppression of ringing and avoiding smoothing. An
equivalent technique is to take the fast Fourier transform (FFT) of the sampled data,
apply a bandpass filter in the Fourier domain, shift in phase and convert back again.
Nearest-neighbour (zero-degree spline) interpolation is the simplest and fastest to com-
pute method. This is equivalent to convolving the sampled data with a box function,
Π(x/a), instead of sinc(x/a). Since the Fourier transform is a sinc function, it does not
actually cut off F (u) at the Nyquist frequency, although it attenuates higher frequen-
cies (Figure 2.24). The first zeros are at ±1/a, or twice the Nyquist frequency. Beyond
that, sinc(au) goes negative and therefore reverses the phase of some frequencies. The
frequencies passed above 1/2a correspond to the ‘blockiness’ of the nearest-neighbour
interpolant, which can be considered a reconstruction error.
In terms of reproducing the analogue telescope image, the nearest-neighbour algorithm
is about the least accurate method. Nevertheless, the values are guaranteed to be rea-
sonable, since they are the same as the original samples: blockiness is an artifact, but a
well-behaved and understood one. It should also be noted that this interpolant does not
smooth the data, as each value comes from a single sample. Resampling with the same
interval at new points effectively just introduces a small coordinate shift. Since both
the values and gradient are discontinuous, the reconstruction cannot be used directly for
fitting smooth contours or velocity fields (although it might be possible to resample first
at the same rate and then use a smooth interpolant, if the sampling grid is regular).
Linear (first-degree spline) interpolation is equivalent to convolving the sampled data
with a triangle, or tent, function, Λ(x/a) (Figure 2.25). This is the self-convolution of the
box function, Π(x/a), and has width 2a. Its Fourier transform is therefore ∝ sinc2 (au),
which dies away much quicker than sinc(au) in the wings and is positive everywhere,
but still does not reach zero until ±1/a. The frequencies passed above the Nyquist limit
Observational procedures and data reduction 65
x
−a a
Figure 2.25. Linear interpolation is equivalent to convolving a series of point samples with
Λ(x/a). In this example, the dark line shows the summation of three individual triangle func-
tions, one per sample.
a) b) c)
Figure 2.26. Resampling a linear interpolant at the original rate is equivalent to sampling in
the first place with pixels of a size and sensitivity distribution that depend on the offset. In
each of the three cases above, the height at each point represents the relative contribution to
the interpolated value at the position of the arrow.
5
Since the sampling rate is unchanged, the larger pixels would have to overlap, as for
multiple dither positions.
66 James E.H. Turner
Fourier transform is therefore ∝ sinc4 (au), which quickly drops to ∼ 0 beyond the central
peak, but also suppresses frequencies below the Nyquist limit substantially. In fact, the
Central Limit Theorem (Bracewell, 2000) asserts that the result of repeated convolution
tends towards a Gaussian curve, so higher-order b-splines and their transforms are all
Gaussian-like. Convolving sampled data with a cubic b-spline gives a well-behaved result,
which is more accurate than nearest-neighbour or linear interpolants but which causes
strong smoothing of the data. It is also not strictly an interpolation because it does not
pass through the original samples.
Better results can be achieved using a cubic spline curve that does interpolate the data.
Instead of simply convolving samples with the cubic b-spline kernel, a set of coefficients
can be derived such that the interpolant behaves more like the ideal sinc reconstructor
(e.g. Lehmann et al., 1999). This is equivalent to convolution with a kernel that is similar
to a sinc function, with both positive and negative values but less ‘rippling’ beyond the
central peak. Its Fourier transform is box-like, but with more of a tapered cut-off. In
practice, cubic spline interpolation provides a good approximation to sinc reconstruction,
but with better cut-off characteristics and a reasonable number of values to compute at
each point. It is therefore the best compromise for many applications involving reasonably
well-sampled data.
Polynomial interpolants can also be constructed that directly approximate the form of
a sinc function, for use as convolution kernels. These can again be simple and relatively
compact, producing sinc-like results with less computation and less oscillation than a
summation of sinc functions. Some other possible basis functions include wavelets, Fourier
series, Taylor series and Gaussian summations.
Irregular sampling
It is possible to formulate more general theories for irregular sampling, but the relatively
advanced mathematics is beyond the scope of this lecture (for an overview, see Marvasti,
2001). In principle, a band-limited signal can be specified by as few samples as it contains
unknown frequency coefficients – at least the average interval between samples must
meet the regular Nyquist criterion (e.g. Beutler, 1966). However, the vectors describing
sample weights in the space of the basis functions are no longer orthogonal for non-
uniform spacing, so reconstructing the signal can become an ill-conditioned problem.
For guaranteed stability, some methods require that the maximum separation between
samples is no greater than the Nyquist interval.
Reconstruction from irregular samples can be approached in more than one way. For
specific sampling patterns, an analytical solution can be derived; a simple case is that
of interlaced grids, dealt with by Bracewell (2000). Alternatively, one can approximate
the signal by fitting regular basis functions, perhaps iteratively (e.g. Strohmer, 1993).
Either way, any clustering of samples can lead to strong amplification of noise in the
reconstruction, especially when working close to the Nyquist limit (e.g. Lauer, 1999).
Irregular sampling is most likely to be of concern where the spacing varies between
adjacent elements at the IFU input, particularly with optical fibres (assuming the sam-
ple density is high enough in the first place to allow interpolation). It may also occur
indirectly in the rare case where fibre profiles are undersampled at the detector (Turner,
2001). The spatial pixels of an image slicer form a grid that appears irregular in two or
more dimensions, but which can be separated into a series of regular grids in 1D. For
datasets that are only slightly non-uniform on large scales due to geometric distortions, it
may be acceptable just to treat the sampling as regular, as long as the Nyquist criterion
is still met, and then resample the interpolant at variable intervals corresponding to a
regular grid in the appropriate system.
Observational procedures and data reduction 67
2.4.5 Summary and application to astronomical data
For well-sampled data, it is possible to resample images with very little degradation, as
if we had sampled with a convenient regular grid to begin with. A diffraction-limited
telescope image can be recovered in its original analogue form as long as we sam-
ple with ≥ 2 pixels across the width of the Airy PSF. For seeing-limited images, at
least 2 pixels/FWHM are needed to reduce aliasing to a level comparable with other
errors, such as flat-fielding systematics or noise. For accurate reconstruction with mini-
mal smoothing, a tapered sinc function or smooth piecewise interpolant such as a cubic
spline is likely to be the best option.
Close to the Nyquist limit or in the presence of image defects, it is best to avoid convo-
lution kernels with too sharp a cut-off in the frequency domain, especially truncated sinc
functions; these accentuate the ringing close to sharp features that occurs when idealiza-
tions such as diffraction-limited resolution are not met. Smooth spline-like interpolants
with compact kernels are better behaved, but still prone to some oscillation. Convolution
kernels that are positive everywhere are insensitive to such issues, but can cause strong
smoothing, especially near the Nyquist limit.
For undersampled data, one only knows the image intensity at the sample points,
with insufficient information to say what happens in between. The ideal strategy is
therefore to avoid interpolating altogether. This is not always practical, however, when
data are on a non-rectangular grid with distortions and perhaps mutual offsets. It may
be necessary to interpolate in order to combine datasets or work with a square grid
(otherwise analysis would require special end-to-end treatment, incorporating basic pro-
cessing). Unfortunately, sinc-like interpolation of undersampled data would introduce
serious aliasing artifacts, whilst smooth, positive convolution kernels could cause exces-
sive loss of detail. The best compromise may therefore be to use simple nearest-neighbour
or linear methods, or the related Drizzle algorithm (Fruchter and Hook, 2002). These are
relatively inaccurate in terms of reconstructing a smooth image, but ensure that values
remain meaningful. They can, however, introduce their own peculiarities in the presence
of image distortions, such as skipping nearest neighbours or variable smoothing.
Whilst the above discussion considers sampling telescope images, the same principles
apply to sampling a dispersed spectrum. The spectral PSF of a slit spectrograph or image
slicer is a rectangular slit (for extended sources) or a narrow section of the telescope PSF,
truncated by the slit (for stars). Although the edges are smoothed slightly by the finite
resolution of the optics, these spectral profiles pass high frequencies more strongly than
a Gaussian or Airy profile of similar width, so 2 pixels/FWHM does not work quite
as well. One may therefore need to be careful to use an interpolant that damps high
frequencies without a sudden cut-off. This is less of an issue for spectra with intrinsically
broad features, where the sampling may be good even though the instrumental PSF for
narrow lines would be undersampled. Fibre or lenslet spectra, on the other hand, often
produce profiles closer to a Gaussian curve, so can be treated more like seeing-limited
images.
a) 5000 30 b) 30
1
4000
3000 20 .75 20
2000 .5
10 10
1000 .25
0 0
200 400 600 800 200 400 600 800
Line (pixels) Line (pixels)
Figure 2.28. Fitting and dividing out the continuum shape in a calibration lamp spectrum a),
to produce a detector pixel flat field b).
towards zero again; it may therefore be more difficult to identify highly saturated pixels
automatically than in the optical.
2.5.5 Flat-fielding
As discussed in Section 2.3, it is necessary to correct for differences in sensitivity both
between detector pixels and across the IFU field. Depending on the instrument, these
aspects can be treated separately or together.
To create a detector flat, a set of several dispersed calibration lamp flats are usually
combined, giving a signal-to-noise ratio high enough that the flat will not contribute
random noise when applied to science data. The bias should, of course, be removed as
well. The spectral continuum shape of the co-added flat may be fitted and divided out at
each spatial position, leaving only small-scale pixel-to-pixel differences (Figure 2.28); any
larger-scale variation in detector sensitivity with wavelength is usually dealt with during
flux calibration, rather than here. The resulting flat should be normalized to an average
70 James E.H. Turner
Figure 2.29. Detector fringing in optical fibre IFU spectra: wavelength increases to the left; the
narrow horizontal stripes are individual fibres; the thicker dark bands are gaps between blocks
of 50 fibres; a small subsection of the detector is shown.
of one, of course, to preserve flux. Raw science spectra are then divided pixel-for-pixel
by the flat, either separately or as part of a combined pixel+IFU correction (see below).
Particularly for CCDs at redder wavelengths, it may be necessary to deal with fringing
due to optical interference (e.g. Figure 2.29). This occurs when light is reflected back
and forth within the detector substrate (or perhaps between filter surfaces) before being
absorbed. For spectroscopy, this is a multiplicative effect; each point on the detector
receives quasi-monochromatic light, so fringing is a function of the total amplitude at
the relevant wavelength. In principle, fringing can therefore be removed by flat-fielding
the detector. However, the process then becomes sensitive to any small drifts in flexure
or grating tilt, since differences in illumination between science exposures and flats alter
the pattern, leaving systematic residuals.
For spectrographs that use a mosaic of separate detectors, any chip-to-chip differences
in flat-fielding must be treated carefully. Colour sensitivity curves can vary between
CCDs, for example, leaving discontinuities that cannot be corrected by scaling each chip
or fitting a smooth lamp continuum across the chips. Instead, the quantum efficiency of
each CCD may need to be factored in as a function of wavelength. Fringing properties
can also vary significantly between chips.
An IFU flat corrects the relative transmissions of different IFU elements, or variations
along image slices, to the same system (Figure 2.30). In the absence of relative flexure
between flats and science exposures, pixel-to-pixel detector variations can also be left
in, giving a combined pixel+IFU flat. This requires a calibration lamp spectrum, since
twilight flats have many sky lines and are thus unusable for calibrating spectral variations.
It is usually possible to match the flexure of science data by taking flats often enough.
This also allows IFU elements to be traced easily by using the flats as a reference. Where
there is flexure between IFU flats and science data, however, both should be divided
by a normalized detector pixel flat, removing super-Nyquist variations, before shifting
them to match. To obtain sufficient signal in between the centres of IFU elements, the
detector flat may need to be taken without the IFU in the beam, perhaps using a long
slit matching the spectral illumination of the detector by the IFU.
Where a calibration lamp does not accurately mimic the illumination of the IFU by the
sky, it can be combined with a twilight exposure to correct the spatial content. The lamp
flat provides detector pixel variations and any differences in wavelength response between
IFU elements, whilst the twilight contributes spatial variations between IFU elements,
or just a large-scale flatness correction to the spatial information from the lamp flat. The
Observational procedures and data reduction 71
Figure 2.30. Example IFU spectra before and after flat-fielding out detector and IFU structure.
From top to bottom, the panels are a) a single image slice, with structure from the slicing optics,
b) the same exposure after a lamp flat is applied, c) a stack of extracted fibre spectra, one per
row and d) the same data after a combined lamp and twilight flat is applied. Twilight spectra
have been used for illustration, since they are bright and highlight structure clearly.
spectral content can be removed from twilight spectra by summing in wavelength, which
also helps with signal-to-noise requirements. For consistent results, it is important to
sum over the same range of wavelength for each element, rather than a fixed number of
pixels. One can then compare the ratios of spatial variations in twilight and lamp flats,
adjusting the latter to match the former.
If IFU elements have differing transmission spectra, the lamp flat should be normal-
ized to account for this, dividing each spectrum by the average over the elements of the
continuum spectrum for the lamp. This preserves spectral corrections to the overall flux
calibration (which is derived from a standard star). It may be necessary to shift and scale
the continuum to match each element. Otherwise, if spectral differences are negligible,
the continuum can be divided out separately from each IFU element, leaving only pixel
variations. One would not normally expect to encounter large wavelength-dependent dif-
ferences across an IFU, but the possibility is worth considering, to avoid extra systematic
errors.
When reducing data from optical fibres and lenslet arrays, one can normally extract
each fibre or lens spectrum to 1D before flat-fielding. The flat can still incorporate pixel
variations, as long as science and flat-field spectra are extracted with the same weights. If
neighbouring lenses or fibres are to be deconvolved in science spectra, the same procedure
should be applied to the flat. Each spectrum is then divided through by the final flat as
normal. An exception is the unusual case of unresolved fibres, which can be treated more
like image slicer data (Turner, 2001). In principle, a possible complication with fibres is
that their relative throughputs could vary with flexure, depending on exactly how each
one is stressed (e.g. Oliveira et al., 2005). In practice, this has not been noticeable in the
author’s experience, but it gives yet another motive for taking flats to match the flexure
of science exposures.
In image slicer data, each slice can be treated much like a slit spectrum, but accounting
for differences in transmission between slices as well as along their lengths. A 2D flat
can be derived for each slice by dividing out the spectral continuum of the lamp from
each image row. To preserve spatial variations, a consistent normalization should be
maintained both along and between slices when doing this.
72 James E.H. Turner
2.5.6 Cosmic ray and bad pixel correction
Before interpolating or co-adding data, any bad pixels should be masked out to avoid
contaminating good values (or generating artifacts). In principle, this can be done either
in 2D or 3D, with the latter providing more information at some expense in simplicity.
Any exposures that are genuinely identical (apart from noise) can simply be co-added
with statistical rejection in 2D. This often applies only for short integrations, however,
since the telescope pointing, airmass and flexure must all remain constant over at least
three frames. An alternative is to use PSF (band limitation) constraints to detect fea-
tures in individual frames that are too narrow or steep to be real (e.g. Rhoads, 2000).
This is more difficult for IFUs that pack information onto the detector with little or no
redundancy, such as image slicers, but easier for smooth and well-sampled fibre or lenslet
PSFs.
Following a similar approach in 3D, one could identify bad pixels in the process
of constructing a datacube. This makes a greater number of neighbours available for
constraining (and interpolating) values, but does not allow for interpolation prior to
3D reconstruction. When co-adding data in 3D, statistical rejection can be applied
to the interpolated values and used alongside data quality arrays from an earlier stage,
to identify bad samples.
Given the small number of spatial elements and high density of information in a typical
IFU dataset, defective pixels may have a relatively high impact. At the same time, they
can be less straightforward to identify than when co-adding direct images, extracting a
2D spectrum of a point source or working with more extended slit spectra. This reduction
stage may therefore require some care.
2.5.7 Instrumental background
To ensure that measurements reflect the true object+sky intensity, any instrumental
background counts must be removed. Especially with fibres, nearby elements at the IFU
input can be widely separated at the detector, so it is possible for varying background
levels to introduce structure on small spatial scales. Where there are gaps between IFU
elements (or blocks of elements) at the output, the background level in the gaps can be
used to model and remove any diffuse scattered light. In the infrared, however, dark cur-
rent and any other target-independent background sources are normally taken care of by
sky subtraction, without special treatment. Note that background light can have differ-
ent colour characteristics from the dispersed object signal and change between mosaiced
detectors.
2.5.8 Spatial calibration
Before data can be transformed into a physical coordinate system, it is necessary to
identify IFU elements (fibres, lenses or slices) on the detector and calibrate their spatial
coordinates. This usually involves locating the spatial elements in a flat-field spectrum
matching the science data, with reference to a known IFU mapping.
For individually resolved optical fibres, one can identify fibre peaks in a cross-section
perpendicular to the dispersion axis (see Figure 2.7) and trace them along the length of
the detector. If individual spectra are heavily blended together, one can use any gaps
between blocks of fibres or perhaps cross-correlation of the fibre throughput pattern to
calculate where individual elements are located. If there is no flexure (e.g. at a Nasmyth
focus), the positions can be pre-determined by masking out groups of fibres in a flat and
tracing one group at a time.
For lenslet IFUs, we are again looking for peaks in cross-section, but with the spectra
distributed around the detector in 2D rather than forming a slit. An optical model is
Observational procedures and data reduction 73
typically used to determine the relative positions and lengths of spectra for a given sky-
to-detector mapping, reducing the problem to one of measuring zero-point offsets from
the peaks.
For image slicers, any gaps between the slices at the output can be used to find their
edges. A special Ronchi screen (consisting of regular parallel lines) may also be placed in
front of the IFU input, to calibrate the alignment and scale of each slice along its length.
If slices are staggered in wavelength, steps in the wavelength solution could provide an
alternative indication as to where the edges are. In the absence of flexure, pre-determined
pixel positions can again be used. Having found the edges of the slices (or Ronchi bands),
the tilt and curvature of each one can be traced along the dispersion axis of the detector.
The usable length of each slice can be delimited according to some data threshold or a
fixed number of pixels from the centre.
2.5.10 Extraction
This stage involves extracting IFU elements from the raw data format to something
convenient for later processing. For fibres or lenses, we usually want to extract each
spatial element to a 1D spectrum, integrating in cross-section with or without weights
(Figure 2.31). This is complicated by the fact that adjacent spectra often overlap substan-
tially in cross-section (e.g. Figure 2.7). For fibres aligned in wavelength, this amounts to
an extra convolution in one spatial dimension: the simplest option is to accept a small loss
of spatial resolution, summing over each fibre along with some signal from its neighbours.
Where spatial resolution is exceptionally important, the alternative is to try deconvolv-
ing neighbouring spectra, at a significant cost in extra noise. Deconvolution is sensitive
74 James E.H. Turner
Figure 2.31. Example of GMOS fibre spectra before and after extraction to 1D by summing
over each fibre in cross-section.
Figure 2.32. Section of a near-infrared image slicer spectrum, before and after sky subtraction.
Many bright pixels with high dark current also subtract out well.
the IFU, one could dither the standard around the field and try extracting multiple
spectra, for calibrating different groups of elements.
2.5.14 Reconstruction in 3D
Reduced data in 1D or 2D arrays can be transformed back to a physical coordinate grid
either before or after scientific analysis, depending on the application. In general, this
requires interpolation. One approach is to resample once onto a 3D datacube in (x, y, λ),
for subsequent analysis and visualization. Special tools (such as the Euro3D software)
can also be used to interpolate ‘on the fly’, as needed during visualization or analysis.
Where applicable, a third possibility is to analyse individual 1D spectra and interpolate
the results onto a 2D map. Finally, one may prefer to perform analysis directly in 3D
at the original, calibrated pixel positions, perhaps by fitting a physical model of the
target. This last method avoids resampling and any associated uncertainties, but makes
it difficult to use existing, general-purpose software.
The best process for transforming data to a uniform grid depends on both the initial
sampling and IFU geometry. For well-sampled datasets, one can interpolate the data
with smooth surfaces that reconstruct the dispersed telescope image accurately, allowing
Observational procedures and data reduction 77
Figure 2.34. a) Example of extracted fibre IFU data from a Seyfert galaxy observation and
b) one image plane of the datacube produced by resampling spatially with smooth, piecewise
polynomials.
resampling onto datacubes or 2D spatial maps with minimal degradation (Figure 2.34).
The result is therefore similar to what a 3D detector might provide more directly. For
undersampled data, extra care must be taken to avoid introducing aliasing artifacts or
excessive smoothing; it is not possible to reconstruct exactly what a 3D detector would
see. One should resample as little as possible – maybe once, to get to 3D – and stick to
positive, compact interpolants, such as nearest-neighbour, linear or Drizzle weightings.
Depending on the geometry of the IFU, one has a choice as to how many dimensions
to interpolate in at once. Resolved fibre or lenslet spectra can be extracted to 1D and
resampled onto a common grid in wavelength before interpolating spatially. Where the
IFU input consists of square lenslets, the spectra can then simply be arranged to form a
2D grid spatially. For lenslets or fibres packed in a triangular (hexagonal) configuration,
a square spatial grid is obtained by interpolating between the spectra at each wavelength
element, in either 1D or 2D. To deal with one dimension at a time, it is possible to
interpolate first along each linear row of hexagons or circles and then across the rows.
With image slices, the (x, λ) coordinates of each 2D spectrum are not trivially separa-
ble, due to image distortions. It therefore makes more sense to interpolate spatially and in
wavelength at the same time. This is the same problem as straightening and wavelength
calibrating a long-slit spectrum in 2D, except that all the slices should be transformed
to the same (x, λ) grid, so cannot be treated completely independently. Once the slices
are rectified, they need just to be stacked in y to get to 3D. Image slicers often have
rectangular spatial pixels, however, with the slice width being different from the detector
pixel size. In those cases, one can interpolate between slices in y, to get to a square spatial
grid.
For any type of IFU data, the alternative to resampling each element individually is
to interpolate processed detector pixel values directly in 3D, after calibrating x, y and λ
for each pixel. This is typically more complicated (except for the simplest interpolants)
than working in lower dimensions, especially where samples are distributed irregularly
in physical coordinates. For interpolating smooth curves, it is significantly easier to find
pre-existing software that works in fewer than three dimensions, on regular grids and
with homogeneous coordinates. For undersampled data, however, transforming straight
to 3D may be the best approach, using something like a nearest-neighbour or Drizzle
scheme to map pixels directly onto a cubic grid. As of 2006, the Gemini IRAF (image
reduction and analysis facility) package includes a task, ‘gemcube’, that can do this using
a version of the Drizzle algorithm in 3D.
78 James E.H. Turner
2.5.15 Atmospheric dispersion
Atmospheric dispersion, or differential refraction, causes the position of a target within
the IFU field to vary with wavelength. This is typically less problematic than for a single
slit, where the target can be shifted entirely out of the aperture at some wavelengths.
Nevertheless, the effect can distort spectra and cause different-sized variations at different
airmasses. Where significant, a correction must therefore be applied before combining
datasets from different airmasses or comparing the positions of spatial features at different
wavelengths.
The spatial offset as a function of wavelength can be determined either using an atmo-
spheric refraction model (given the airmass, position angle and other parameters) or by
centroiding on a point source within the IFU field at different wavelengths. Some details
are discussed in Arribas (1999). After constructing a datacube, different corrective shifts
can then be applied to each wavelength plane. Alternatively, the sky coordinates of each
pixel (as calculated during spatial calibration) can be recalculated before transforming to
3D or comparing with a model. At low airmass or infrared wavelengths without adaptive
optics, it may be safe to assume dispersion is negligible. Also, when integrating flux over
a spatial aperture centred on a point source, dispersion is taken care of without special
treatment.
20
10
Arcsec
0
−10
−20
−20 −10 0 10 20
Arcsec
Figure 2.35. A spatial Voronoi tessellation of IFU data, using bins of approximately constant
signal-to-noise ratio (Cappellari and Copin, 2003).
As a final note, it is worth emphasizing that, with the correct treatment, one can often
use IFS data for high-fidelity imaging, as well as spectroscopy.
Figure 2.36. Example reduction sequence for scientific exposures from the GMOS optical fibre
IFU, using Gemini IRAF package tasks.
is processed in parallel with the scientific image, to reflect how the errors change. When
adding two images, for example, the corresponding errors are added in quadrature.
Sources of error
A process where discrete values vary statistically around a well-defined mean, such as
counting photons, is described by a Poisson distribution:
µn
Pµ (n) = e−µ , (2.12)
n!
Observational procedures and data reduction 81
Figure 2.37. Example reduction sequence for scientific exposures from the GNIRS infrared
image slicer, using Gemini IRAF package tasks. Telluric and flux correction are not shown, but
in this case would be performed between nstransform and nfcube.
with a mean (expected number of photons) of n = µ. The standard deviation from the
√
mean is simply σ ≡ µ, so when counting photons (really photo-electrons), the statistical
error is the square root of the expected number of photons (electrons). In practice, the
error is estimated using the square root of the measured number of electrons, since that
√
is what we know. For large µ, the Poisson distribution is Gaussian, with σ = µ.
The main random sources of measurement error in data are read noise from the detector
electronics, Poisson noise from the science target and sky (or other background) and
Poisson noise from detector dark current. The last two of these can be estimated as
discussed above, directly from the pixel values, whilst read noise is a fixed property of
the instrument. We also have systematic errors – notably inaccuracies in flat-fielding –
introduced during processing or by instrumental variations. These are usually present at
the level of a few percent and are difficult to reduce to zero. Such effects can be more
difficult to account for than noise, but are typically dominated by statistical errors.
For more complex operations on science data, represented by some arbitrary function,
n = f (n1 ), the first derivative of f can be used to estimate how output values vary with
small changes in the input for a given n:
σ 2 = f ′ (n1 )2 σ12 . (2.17)
At the end of the complete reduction process, error propagation allows the final noise
values to be determined relatively accurately just by taking the square root of the final
variance array.
Resampling
In raw data, each pixel has an independent statistical error. If resampling causes smooth-
ing, however, the errors of different pixels become correlated. As an example, linear inter-
polation at the mid-point between samples √ is an averaging of pairs. The error of each
interpolated value is therefore reduced by 2, without halving the number of pixels.
Observational procedures and data reduction 83
Clearly the average signal-to-noise ratio over all pixels cannot have increased, however,
nor
√ can summing two adjacent resampled pixels reduce the error by a further factor of
2, since only three of the original values are involved. What has happened is that the
errors of adjacent pixels are no longer independent.
In a case like this, one could track the errors by propagating a higher-dimensional
covariance matrix, where the covariance of two variables is defined as the expected value
of the product of deviations from the means:
σxy = ⟨(x − x̄)(y − ȳ)⟩ = ⟨xy⟩ − x̄ȳ. (2.18)
The variance of a single variable is therefore equivalent to its own covariance:
σx2 = σxx = ⟨x2 ⟩ − ⟨x⟩2 . (2.19)
,
With this information, the error on a sum ni=1 xi is calculated using
n
$ $
σ2 = σx2 i + 2 σx i x j . (2.20)
i=1 i< j
For more information, see an error analysis or statistics textbook, such as Taylor (1997).
Since this level of error propagation is computationally complex, reduction software
usually does not track covariance. It is nevertheless important to be aware that simple
variance estimates may not be exactly correct after resampling. In the ideal limit of
resampling a sinc interpolant at the original rate, however, covariance is not an issue
because no smoothing occurs (at any given point, the RMS distance of positively and
negatively weighted sample contributions is zero).
! one or more binary tables, using named columns (e.g. XCOORD) and mixed types
(e.g. OBJECT = ‘NGC1068’ or EXPTIME = 120);
Figure 2.38. Example multi-extension FITS structure for part-processed image slicer data,
from the edited output of the IRAF task fxheader. The full file continues until EXTVER=21,
corresponding to the 21st image slice. The ‘MDF’ (mask definition file) extension is a binary
table containing the IFU sky-to-detector mapping. Fibre or lenslet data are more likely to be
stored in a 2D or 3D array in a single EXTVER.
Multiple datasets within a FITS file are kept in separate ‘extensions’. First is a primary
header, containing keywords applicable to the whole file, such as the object name, tele-
scope pointing, airmass, filter and central wavelength. This is followed by one or more
numbered and/or named extensions, each containing an image array or binary table,
along with any header keywords specific to that particular dataset. Figure 2.38 shows an
example FITS structure for image slicer data, including variance and data quality arrays
and a table with details of the IFU mapping.
D A
T
T
A
A
Q
ST
D
(Potential) vector
Example of valid data
ID
Sp #
C o d1
G 2
el
up
d
or
or
ax
ro
Co
D
ti h
le c I
ex
ar gt
Se trum
NS tion
St le n
nd
Sp X
PA
ec
ec
Sp
Figure 2.39. The table structure of the Euro3D data format (From Kissler-Patig et al., 2003).
visualization tool (although the latter can also read datacubes). Both 1D or 2D spectra
and the attributes describing them are stored together in a binary table (Figure 2.39).
The format is closer to the raw data than a cube, avoiding resampling until needed during
visualization or analysis. The file size is therefore minimal, as for row-stacked spectra.
One limitation is that special libraries are needed to work with the format, although
these are now publicly available in the C programming language from the Centre de
Recherche Astronomique de Lyon.
2.7 Acknowledgements
Thanks to the organizers for their kind invitation to the Winter School. My initial studies
of IFU data reduction and sampling were aided by input from Jeremy Allington-Smith,
Robert Content and the instrumentation group at Durham, UK. The Gemini Observatory
has since given me the opportunity to commission and observe with numerous IFUs. Some
of the details in these lectures are based on my work or discussions with Bryan Miller at
Gemini, Frank Valdes at NOAO, the Euro3D team and other colleagues.
This work was supported by the Gemini Observatory, which is operated by the Associa-
tion of Universities for Research in Astronomy, Inc., on behalf of the international Gemini
partnership of Argentina, Australia, Brazil, Canada, Chile, the UK and the USA.
86 James E.H. Turner
REFERENCES
Allington-Smith, J.R., Content, R. (1998), PASP, 110, 1216
Allington-Smith, J.R., Murray, G., Content, R. ET AL. (2002), PASP, 114, 892
Allington-Smith, J.R., Dubbeldam, C.M., Content, R. ET AL. (2007), MNRAS, 376, 785
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B. ET AL. (1999), A&AS, 136, 189
Bacon, R., Copin, Y., Monnet, G. ET AL. (2001), MNRAS, 326, 23
Bershady, M.A., Andersen, D.R., Harker, J. ET AL. (2004), PASP, 116, 565
Beutler, F.J. (1966), SIAMR, 8, 328
Bowen, I.S. (1938), ApJ, 88, 113
Bracewell, R.N. (2000), The Fourier Transform and Its Applications, McGraw-Hill, 3
Cappellari, M., Copin, Y. (2003), MNRAS, 342, 345
Cardiel, N., Gorgas, J., Cenarro, J., Gonzalez, J.J. (1998), A&AS, 127, 597
Fruchter, A.S., Hook, R.N. (2002), PASP, 114, 144
Goodman, J.W. (1968), Introduction to Fourier Optics, McGraw-Hill
Hecht, E. (1987), Optics, Addison-Wesley, 2
Kissler-Patig, M., Copin, Y., Ferruit, P., Pécontal-Rousset, A., Roth, M.M. (2003),
Euro3D data format, format definition, http://www.aip.de/Euro3D
Kotel’nikov, V.A. (1933), On the transmission capacity of ‘ether’ and wire in electrocommu-
nications, Izd. Red. Upr. Svgazi RKAA
Lauer, T.R. (1999), PASP, 111, 227
Lehmann, T.M., Gönner, C., Spitzer, K. (1999), Survey: interpolation methods in medical
image processing, IEEE Trans. Medical Imaging, 18, 1049
Marvasti, F.A. ed. (2001), Nonuniform sampling: theory and practice, Kluwer Academic/
Plenum
Moffat, A.F.J. (1969), A&A, 3, 455
Nyquist, H. (1928), Certain topics in telegraph transmission theory, AIEE Trans, 47, 617
Oke, J.B. (1990), AJ, 99, 1621
Oliveira, A.C., de Oliveira, L.S., dos Santos, J.B. (2005), MNRAS, 356, 1079
Pasquini, L., Avila, G., Blecha, A. ET AL. (2002), Installation and commissioning of
FLAMES, the VLT Multifibre Facility, Msngr, 110, 1
Raab, W., Poglitsch, A., Looney, L.W. (2004), FIFI LS: the far-infrared integral field
spectrometer for SOFIA, SPIE, 5492, 1074
Rhoads, J.E. (2000), PASP, 112, 703
Shannon, C.E. (1949), Communication in the presence of noise, IRE Proc., 37, 10
Strohmer, T. (1993), Efficient methods for digital signal and image reconstruction from nonuni-
form samples, PhD, University of Vienna
Taylor, J.R. (1997), An introduction to error analysis, the study of uncertainties in physical
measurements, University Science Books, 2
Turner, J.E.H. (2001), Astronomy with integral field spectroscopy: observation, data analysis
and results, PhD, University of Durham, UK
Vanderriest, C.A. (1980), PASP, 92, 858
Whittaker, E.T. (1915), On the functions which are represented by the expansions of the
interpolation theory, Proceedings of the Royal Society Edinburgh, Section A, 35, 181
Wright, G.S., Rieke, G.H., Colina, L. ET AL. (2004), The JWST MIRI instrument concept,
SPIE, 5487, 653
3. 3D spectroscopic instrumentation
MATTHEW A. BERSHADY
In this chapter we review the challenges of, and opportunities for, 3D spectroscopy and
how these have led to new and different approaches to sampling astronomical infor-
mation. We describe and categorize existing instruments on 4 m and 10 m telescopes.
Our primary focus is on grating-dispersed spectrographs. We discuss how to optimize
dispersive elements, such as VPH gratings, to achieve adequate spectral resolution,
high throughput, and efficient data packing to maximize spatial sampling for 3D spec-
troscopy. We review and compare the various coupling methods that make these spec-
trographs ‘3D’, including fibres, lenslets, slicers, and filtered multi slits. We also describe
Fabry–Perot (FP) and spatial-heterodyne interferometers, pointing out their advantages
as field-widened systems relative to conventional, grating-dispersed spectrographs. We
explore the parameter space all these instruments sample, highlighting regimes open
for exploitation. Present instruments provide a foil for future development. We give an
overview of plans for such future instruments on today’s large telescopes, in space and in
the coming era of extremely large telescopes. Currently-planned instruments open new
domains but also leave significant areas of parameter space vacant, beckoning further
development.
87
88 Matthew A. Bershady
Figure 3.1. Sampling the datacube with equal volumes and detector elements.
! spatial resolution dΩ, i.e. the sampling element on the sky (fibre, lenslet, slicer slitlet,
R
where ∆λ is the sampled spectral range, and £ is the cost in the suitable local currency.
To this figure of merit one may add the product Rn × dΩm , where n, m = 1 if resolution is
science-critical in the spectral and spatial domains (respectively), n, m = −1 if coverage
is science-critical, or n, m = 0 if resolution and coverage are science-neutral (in which
case you are not trying hard enough!).
From this discussion it is clear that a suitable choice of merit function is complicated,
and must be science driven. The relative evaluation of instruments cannot be done sensi-
bly in the absence of a science-formulated FOM; the outcome of any sensible evaluation
will therefore depend on the science formulation. For this reason, when we compare
instruments we strategically retreat and explore the multi-dimensional space of the fun-
damental parameters of spatial resolution, spectral resolution, specific grasp, total grasp,
spectral power, and NR versus NΩ .
Figure 3.2. Night sky near 0.8 µm at 250 < R < 33,000.
change in the distribution of background levels. In Figure 3.3, the normalized, cumu-
lative distribution of resolution elements as a function of background level is plotted
for different instrumental resolutions. For background-limited measurements, the S/N is
proportional to the inverse square root of the background level. Hence the median back-
ground level gives an effective scaling for sensitivity gains with spectral resolution. It
can be seen that the largest changes in the median background level occur between
1000 < R < 4000, but significant gains continue at higher resolution. The result can
be qualitatively generalized to other wavelengths in the 0.7–2.2 µm regime. While the
lines become more intense moving to longer wavelengths, the power spectrum (in wave-
length) of the lines appears roughly independent of wavelength in this regime (compare
3D spectroscopy instrumentation 91
Figure 3.4. Maximum spectral resolution versus telescope diameter to stay background (versus
detector) limited for different assumptions of instrument efficiency (ϵ) and spatial sampling (dΩ).
The solid line assumes ϵ = 0.15 and dΩ = 1 arcsec2 .
Maihara et al., 1993; Hanuschik, 2003). Note this is a qualitative assessment that should
be formally quantified.
α
w D1
grating size = Wg
Figure 3.5. Basic spectrograph layout schematic for reflective/refractive collimator, reflection
grating and refractive camera.
overlooked, and also the field of instrumentation advances rapidly. Reports of additional
instruments or corrections are welcome.
1
A field lens, which sits near a focus to avoid introducing power into the beam, serves to
move the spatial pupil to a desirable location in the system. This is often the grating, but in
general can be the location such that the overall system vignetting is minimized. A white-pupil
design (Baranne, 1972; Tull et al., 1995) is one that re-images a pupil placed on a grating,
typically on to a second grating (e.g. a cross-disperser) or the camera objective. It is ‘white’
because the pupil image location is independent of wavelength even though the light is dispersed.
94 Matthew A. Bershady
Reflection gratings
Reflection gratings come in three primary varieties: ruled surface-relief (SR), holograph-
ically etched SR, or volume-phase holographic (VPH). We list the pros and cons of each
of these.
Ruled SR gratings have the advantage of control over the groove shape, blaze and
density, which provides good efficiency in higher orders (e.g. echelle) at high dispersion.
There are existing samples of masters with replicas giving up to 70% efficiency, but 50–
60% efficiency is typical, with 40% as coatings degrade. Scattered light and ruling errors
can be significant, and existing masters are limited in type and size. It does not appear
to be possible to make larger masters with high quality at any reasonable cost.
Holographically etched SR gratings have low scattered light, the capability to achieve
high line density (hence high dispersion) and large size. However, they have low efficiency
(< 50%) because symmetric grooves put equal power in positive and negative orders.
VPH gratings can be made to diffract in reflection (Barden et al., 2000), but have
not yet been well developed for astronomical use. Reflection gratings can be coupled to
prisms to significantly enhance resolution via anamorphing (Wynne, 1991).
Transmission gratings
Transmission gratings are either SR or VPH and, when coupled with prisms, are referred
to as grisms. SR transmission gratings and grisms are efficient at small angles and low
line densities (good for low-resolution spectroscopy), but are inefficient at large angles
and line densities due to groove shadowing. Transmission echelles do exist, but have 30%
diffraction efficiencies or less.
VPH gratings and grisms are virtually a panacea. They are efficient over a broad range
of line densities and angles. Any individual grating is also efficient over a broad range of
angles (what is known as a broad ‘superblaze’; see below). Peak efficiencies are as high as
90%; they are relatively inexpensive to make and likewise to customize; and they can be
made to be very large (as large as your substrate and recording beam, now approaching
0.5 m). Their only disadvantages are that they have, to date, been designed for Littrow
configurations.
It is worth dwelling somewhat on the theory and subsequent potential of VPH grat-
ings. There still remain manufacturing issues of obtaining good uniformity over large
areas (Tamura et al., 2005), but it is reasonable to be optimistic that refinement of the
process will continue at a rapid pace. Application in the NIR for cryogenic systems is
also promising: coefficient thermal expansion mismatch between substrate and diffracting
gelatin, potentially causing delamination, does not appear to be a concern (W. Brown,
private communication, this Winter School). Blais-Ouellette et al. (2004) have confirmed
that diffraction efficiency holds up remarkably well at 77 K, but that the effective line den-
sity changes with thermal contraction. We can expect most grating-fed spectrographs in
the future will use VPH gratings alone or in combination with conventional (e.g. echelle)
gratings. The capabilities of VPH gratings will open up new design opportunities, many
of which will be well suited to 3D spectroscopy.
R
A
Figure 3.6. Resolution merit function and anamorphic factor for blazed VPH gratings with
mean gel index n2 = 1.43. Typical SR gratings have 1.05 < 1/r < 1.2.
The anamorphic factor and dispersion are still defined in terms of α and β as given in
Section 3.2.1. With the interrelation of these angles as given above, it is easy to show
the logarithmic angular dispersion at the Bragg wavelength is:
% ) * &-
sin α
dβ/dlogλ = 2n2 cos φ sin sin −1
−φ cos β. (3.6)
n2
Figure 3.7. Novel grating modes. a) Conventional broad-band application now becoming a
staple of modern spectrographs. b) Double-grating geometry yielding a net dispersion of ∼ the
sum of the two individual grating dispersions (gratings are not necessarily identical, but angles
must be adjusted accordingly). c) Double-grating geometry yielding a narrow-band filter with
field-dependent bandpass given by the Bragg condition (gratings are identical) d)–f) Narrow
bandpass gratings unblazed (d), blazed (e), and combined (f) to form a notch grating. Other
modes are discussed in the text. Panels a), d)–f) show both the grating configuration as well as
a cartoon-sketch of the diffraction efficiency as a function of location of the detector dispersion
axis, labelled for the mean wavelength regime of the diffraction bandpass.
merit function at α = 60, but a 51% gain in the anamorphic demagnification. With suit-
ably good optics and detector sampling the demagnified image, this equates directly into
an increase in the number of independent spectral resolution elements, replete with a 72%
increase in spectral coverage. The loss in resolution can easily be made up by slightly
increasing α (in this case, from 60 to 63 deg) and modulating Λ in the grating design
to tune the wavelength. Instruments with blazed, high-angle VPH gratings with tilts of
5 < φ < 15 deg will allow for the high resolution needed to work between sky lines, while
efficiently packing spectral elements on to the detector. This is critical in the context of
3D spectroscopy, where room must also be made for copious spatial elements.
2
A true cross-dispersed echelle-like grating would work, in principle, with two layers, rotated
by 90 degrees. VPH gratings have not yet been made with high efficiency in multiple orders, but
see Barden et al. (2000) for measurements up to order 5.
3D spectroscopy instrumentation 99
3.2.6 Direct fibre coupling
The simplest and oldest of methods consists of a glued bundle of bare fibres mapping
the telescope to spectrograph focal surfaces. With properly doped, AR-coated fibres
throughput can be at or above 95%, which can be compared to 92% reflectivity off one
freshly coated aluminium surface. These have the distinct advantage of low cost and high
throughput. As with all fibre-based coupling, there is a high degree of flexibility in terms
of reformatting the telescope to spectrograph focal surfaces (for example, it is easy to mix
sky and object fibres along slit), and the feeds can be integrated into existing long-slit,
multi-object spectrographs. However, bare fibre integral field units (IFUs) are not truly
integral and do not achieve higher than 60–65% fill factors (on the deleterious effects
of buffer stripping of small fibres, see Oliveria et al., 2005). This coupling is perhaps
the most cost-effective mode for cases where near-integral sampling is satisfactory, and
preservation of spatial information is not at a premium.
Telecentricity
Because azimuthal scrambling symmetrizes a beam, if the input light-cone is misaligned
with the fibre axis, the output beam (f -ratio) is faster. This is not FRD. To avoid this
effect, fibre telecentric alignments of under a degree are needed even for f -ratios as fast
as 4–6 (Bershady et al., 2004; Wynne and Worswick, 1989).
Causes of FRD
Excessive FRD in fibres is due to stress. Hectospec (Fabricant et al., 2005) embodies
an excellent example of how to properly treat fibres and fibre cabling (Fabricant et al.,
1998; see also Avila et al., 2003 in the context of FLAMES on VLT). Fibre termination
100 Matthew A. Bershady
Figure 3.8. Output fibre irradiance (encircled energy versus beam f -ratio) for fibre cables on
the WIYN Bench Spectrograph. The input beam profile is an unapodized f /6.3 beam with an
f /17 central obscuration (labelled). Output beam profiles are faster, due to FRD, and are well
fitted by a Sersic model of index 1/n = 5 (S. Crawford, private communication).
and polishing can also induce stress. Bershady et al. (2004) discuss some other IFU-
related issues in terms of buffering. However, even for perfectly handled fibres, there is
internal scattering, the cause of which has long been a debate. Nelson (1988) suggested
a combination of (i) Rayleigh scattering (variation in fibre refractive index); (ii) Mie
scattering (fibre inhomogeneities comparable to the wavelength); (iii) stimulated Raman
and Brillouin scattering (not relevant at low signal level in astronomical applications);
and (iv) micro-bending. Micro-bending seems a likely culprit; it is the unsubstantiated
favourite in the literature. Micro-bending models predict a wavelength-dependent FRD.
While Carrasco and Parry (1994) tentatively see such an effect, neither Schmoll et al.
(2003) nor Bershady et al. (2004) confirm the result. However, these studies use different
measurements methods. More work is required to understand the physical cause(s) of
FRD, and with this understanding, perhaps, reduce the amplitude of the effect. We find
FRD produces an output fibre beam profile which can be well-modelled by a Sersic
function (Figure 3.8; S. Crawford, private communication). This says something about
either the scattering model or how seriously to take physical interpretations of Sersic-law
profiles of galaxies.
in the spectral dimension is high due to large anamorphic factors. Other systems have
significant spectral overlap. For example, staggered slits, where fibres are separated by
only their active diameter (COHSI; Kenworthy et al., 1998) make it difficult to extract a
clean spectrum and optimize S/N at the same time, but the spatial multiplex is increased.
There is no one right answer, but definitely a decision worthy of science-based consid-
eration.
Summary of instruments
Some of the first IFUs were on the KPNO 4 m RC spectrograph: DensePak-1 followed by
DensePak-2 (Barden and Wade, 1988; for other early IFUs, see also Guerin and Felenbok,
1988). The last incarnation (Barden et al., 1998) was on WIYN. Conceptually, these
instruments spawned SparsePak (WIYN; Bershady et al., 2004) and PPak (PMAS, Calar
Alto; Verheijen et al., 2004; Kelz et al., 2006). A more versatile single instrument suite,
built for the William Herschel Telescope (WHT), is INTEGRAL (WYFOS, Wide-Field
Optic Spectrograph), which offers several plate-scales and formats (Arribas et al., 1998),
and a sophisticated and well-thought-out mapping between telescope and spectrograph
focal planes. These are all shown in Figure 3.10. GOHSS is one case of a NIR (0.9–1.8 µm)
application (Lorenzetti et al., 2003). VIRUS (Hill et al., 2004) and APOGEE (Allende
Prieto et al., 2008) are the only planned future instruments. A summary of existing and
future optical and NIR direct fibre-fed IFU spectrographs are listed in Table 3.1.
DT Ω dΩ
Instrument Tel. (m) (arcsec2 ) NΩ ∆λ/λ R NR ϵ
VIRUS HET 9.2 32604 1.0 32604 0.505 811. 410 0.16
APOGEE SDSS 2.1 942 3.14 300 0.10 23500 2350 0.22
Figure 3.10. Direct-fibre IFUs on optical spectrographs. The top row shows the legacy started
by S. Barden with DensePak-1 and DensePak, leading to SparsePak, PPaK on the KPNO 4 m,
WIYN, and Calar Alto, respectively. The bottom row shows Hexaflex and INTEGRAL on WHT
with their multiple, selectable bundles and ample sky-fibres.
3D spectroscopy instrumentation 103
At the output stage, the option exists to reform the (now azimuthally scrambled) slit
image with an output micro-lens linear array, or to use bare fibres. Without lenslets,
the input f -ratio to the spectrograph will be faster, which means there is less possibility
for geometric demagnification via a substantially faster camera. In this case the spec-
trograph also re-images the fibre-scrambled telescope pupil: the image varies with tele-
scope illumination, while the ray bundle distribution (far-field) varies with the telescope
image.
The positive attributes of lenslet-fed fibre arrays are: (i) improved filling factors to
near unity; and (ii) control of input and output fibre f -ratio. The latter permits effective
coupling of a slow telescope f -ratio to fibre input at a fast, non-lossy beam speed and like-
wise permits effective coupling of fibre output to spectrograph. The negative attributes of
this coupling method include (iii) increased scattered light (from the lenslet array); (iv)
lower throughput (due to surface reflection, scattering, and misalignment). For example,
typical lenslet + fibre units yield only 60–70% throughput (Allington-Smith et al., 2002).
When there is a science premium on truly integral field sampling, the above two factors
do not outweigh the filling factor improvements. Finally, there is the more subtle effect of
whether or not to use output lenslets. Aside from the matter of f -ratio coupling, there is
the issue of whether swapping the near- and far-field patterns is desirable for controlling
systematics in the spectral image. It amounts to assessing whether the spectrograph is
‘seeing-limited’, i.e. limited by spatial changes in the light distribution within the slit
image formed by the fibre and lenslet, or aberration limited.
Prime examples of optical instruments on 8 m-class telescopes include VIMOS
(Le Fevre et al., 2003), GMOS (Gemini-N,S, Allington-Smith et al., 2002), and
FLAMES/GIRAFFE in ARGUS or multi-object IFU modes (Avila et al., 2003).3 Typ-
ical characteristics of these devices are fine spatial sampling (well under an arcsec) and
modest spectral resolution. ARGUS is an exception, achieving resolutions as high as
39,000. Its multi-object mode is also unique, and powerful (see later discussion). On 4–
6 m-class telescopes there are PMAS (Roth et al., 2005), Spiral+AAOmega (Kenworthy
et al., 2001; Saunders et al., 2004), MPFS (Afanasiev et al., 1990)4 and IMACS-IFU
(Schmoll et al., 2004).5 Compared to most direct-fibre IFUs on comparable telescopes,
these instruments also have finer spatial sampling.
NIR instruments include SMIRFS (Haynes et al., 1999), and COHSI, which is a pre-
cursor – in some regards – to CIRPASS (Parry et al., 2004). An interesting application
of flared fibres is discussed by Thatte et al. (2000) for cryogenic systems.
A summary of existing and future optical and NIR lenslet + fibre coupled IFU spec-
trographs is listed in Table 3.2. While it may seem surprising that no future instruments
appear to be planned, we will discuss one possible instrument for the 30 m Telescope
(TMT) below.
3
See also www.eso.org/instruments/flames/inst/Giraffe.html.
4
See also www.sao.ru/hq/lsfvo/devices/mpfs/mpfs main.html.
5
See also www.ociw.edu/instrumentation/imacs.
104 Matthew A. Bershady
Tel. DT Ω dΩ
Instrument Method (m) (arcsec2 ) NΩ ∆λ/λ R NR ϵ
PMAS Calar Alto 3.5 64. 0.5 256 0.11 9400 1000 0.15
3.5 64. 0.5 256 0.52 1930 1000 0.15
3.5 144. 0.75 256 0.11 9400 1000 0.15
3.5 144. 0.75 256 0.52 1930 1000 0.15
3.5 256. 1.0 256 0.11 9400 1000 0.15
3.5 256. 1.0 256 0.52 1930 1000 0.15
SPIRAL AAT 3.9 251. 0.49 512 0.29 1700 495 0.25
3.9 251. 0.49 512 0.07 7500 495 0.25
MPFS SAO 6.0 256. 1.0 256 0.12 8800 1024 0.045
6.0 64. 0.25 256 0.47 2200 1024 0.045
IMACS-IFU Magellan 6.5 62.0 0.031 2000 0.61 2500 4096 0.19
6.5 37.7 0.031 1200 0.31 7500 2340 0.17
GMOS Gemini 8.0 49.6 0.04 1500 0.21 3450 730 ···
8.0 49.6 0.04 1500 0.32 2300 730 ···
8.0 49.6 0.04 1500 0.82 890 730 ···
8.0 24.8 0.04 750 0.42 3450 1460 ···
8.0 49.6 0.04 1500 0.64 2300 1460 ···
8.0 49.6 0.04 1500 1.00 890 1460 ···
VIMOS VLT 8.0 2916. 0.45 6400 0.6 250 150 ···
8.0 698. 0.11 6400 0.6 250 150 ···
8.0 729. 0.45 1600 0.2 2500 500 ···
8.0 174.5 0.11 1600 0.2 2500 500 ···
ARGUS/IFU VLT 8.0 83.9 0.27 315 0.105 11000 1155 ···
8.0 83.9 0.27 315 0.042 39000 1625 ···
ARGUS VLT 8.0 27.7 0.09 315 0.105 11000 1155 ···
8.0 27.7 0.09 315 0.042 39000 1625 ···
none
COHSI UKIRT 3.8 8.5 0.85 100 0.26 500. 128 ···
SMIRFS UKIRT 3.8 24.2 0.34 72 0.023 5500. 128 ···
CIRPASS Gemini 8.0 54.5 0.13 490 0.41 2500. 1024 ···
8.0 54.5 0.13 490 0.085 12000. 1024 ···
8.0 27.0 0.06 490 0.41 2500. 1024 ···
8.0 27.0 0.06 490 0.085 12000. 1024 ···
none
deflections, the slices are re-aligned end to end as in a long slit, which then feeds a
conventional spectrograph.
The latest incarnation is the so-called ‘Advanced Image Slicer’ (AIS) concept: a three-
element system, introduced and nicely illustrated by Allington-Smith et al. (2004). In
short, the slicer mirrors at the telescope focal plane divide it into strips and have power
to place the telescope pupil on the next slicer element. This is desirable to keep these
elements small and the slicer compact. The second element is an array of pupil mirrors
3D spectroscopy instrumentation 105
DT Ω dΩ
Instrument Tel. (m) (arcsec2 ) NΩ ∆λ/λ R NR ϵ
UIST UKIRT 3.8 19.8 0.06 344 0.15 3500 512 ···
PIFS Palomar 5.0 51.8 0.45 115 0.23 550 128 0.22
5.0 51.8 0.45 115 0.10 1300 128 0.22
NIFSa Gemini 8.0 9.0 0.01 900 0.19 5300 1007 ···
GNIRSa Gemini 8.0 15.4 0.023 684 0.301 1700 512 ···
8.0 15.4 0.023 684 0.087 5900 512 ···
SPIFFI VLT 8.0 0.54 0.006 1024 0.34 3000 1024 0.3
8.0 10.2 0.001 1024 0.34 3000 1024 0.3
8.0 64.0 0.06 1024 0.34 3000 1024 0.3
KMOSa VLT 8.0 188.0 0.04 4204 0.28 3600 1000 ···
FISICAa GTC 10.4 72.0 0.53 136 0.79 1300 1024 ···
a
Advanced image slicer design.
(one per slice), which reformat the slices into a pseudo-slit, where they form an image
of the sky. A tertiary field lens (a lenslet for each slice) controls the location of the
pupil stop in the spectrograph. This is critical for efficient use of the spectrograph. All-
mirror designs exist for the NIR (FISICA, Eikenberry, 2004b), taking advantage of lower
scattering at longer wavelengths to machine monolithic elements. Catadioptric designs
exist for the optical (MUSE, Henault et al., 2004). Here the pupil lenses replace pupil
mirrors, which aids the geometric layout of the spectrograph system.
The salient features of image slicers are (i) they are the only IFU mode to preserve
all spatial information. All other coupling modes destroy spatial information within the
sampling element, either by fibre scrambling or pupil imaging (below). (ii) Image slicers
are also the most compact at reformatting the focal plane on to the detector. (iii) They
can be used in cryogenic systems and at long wavelengths where fibres do not transmit
(although lenslet arrays also accomplish this; see next section). There are some disad-
vantages, including (iv) scattered light from the slicing mirrors (diamond-turned optics
cannot be used in the optical), and (v) a lack of reformatting freedom. The latter is
perhaps less of a concern given that the image is being preserved. However, for possible
multi-object modes, particular attention must be paid to the design of the required relay
optics to avoid efficiency losses.
We summarize the existing and planned instruments in Table 3.3. The length of the
list, particularly in the planned instruments, marks a sea change over the last few years
106 Matthew A. Bershady
away from fibre + lenslet coupling. While slicers originated for NIR instruments – starting
with the now-defunct MPE-3D (Thatte et al., 1994) – the list of planned optical slicers
is extensive. Existing NIR instruments include PIFS (Palomar Integral Field Spectro-
graph; Murphy et al., 1999) and UIST (UKIRT Imager-Spectrometer; Ramsay Howat
et al., 2006)6 on 4 m-class telescopes; NIFS (Near-Infrared Integral Field Spectrograph;
McGregor et al., 2003), GNIRS (Gemini Near Infrared Spectrograph; Allington-Smith
et al., 2004) and SPIFFI (Spectrometer for Infrared Faint Field Imaging; Eisen-
hauer et al., 2003; Iserlohe et al., 2004), on 8 m-class telescopes. SINFONI7 on VLT
(Bonnet et al., 2004) in particular has shown the power of NIR adaptive optics (AO)
coupled to an image slicer at moderate spectral resolution achieving 20–30% throughput.
Future NIR instruments include KMOS (Sharples et al., 2004) – a multi-object system
discussed below – and FISICA. Below we also discuss three planned NIR instruments for
space.
While the only existing optical instrument is ESI (Echellette Spectrograph and Imager;
Sheinis et al., 2002, 2006), future optical instruments include WiFeS (Wide-Field Spec-
trograph; Dopita et al., 2004), SWIFT (Short Wavelength Integral Field Spectrograph;
Goodsall et al., this workshop) and MUSE (Multi-Unit Spectroscopic Explorer; Bacon
et al., 2004 and references therein). ESI is unique in being the only cross-dispersed IFU
system. While the number of spatial elements is modest, ESI has enormous spectral mul-
tiplex (at medium spectral resolution and good efficiency), the largest of any instrument
planned or in existence.
6
See www.jach.hawaii.edu/UKIRT/instruments/uist/uist.html for sensitivities.
7
SPIFFI + MACAO; Spectrograph for Integral Field Observations in the Near Infrared =
Spectrograph for Infrared Faint Field Imaging + Multi-Application Curvature Adaptive Optics.
3D spectroscopy instrumentation 107
DT Ω dΩ
Instrument Tel. (m) (arcsec2 ) NΩ ∆λ/λ R NR ϵ
SAURON WHT 4.2 1353 0.88 1577 0.11 1213. 128 0.147
4.2 99 0.07 1577 0.10 1475. 150 0.147
OASIS WHT 4.2 1.92 0.002 1100 0.50 1000. 400 ···
4.2 31.0 0.026 1100 0.50 1000. 400 ···
4.2 180. 0.17 1100 0.50 1000. 400 ···
none
OSIRIS Keck 10.4 1.2 0.02 3000 0.12 3400. 400 ···
10.4 30. 0.10 3000 0.12 3400. 400 ···
10.4 0.3 0.02 1019 0.47 3400. 1600 ···
10.4 7.5 0.10 1019 0.47 3400. 1600 ···
none
i.e. there is limited spectral coverage at a given resolution. Spectral extraction is critical
to minimize crosstalk while maximizing S/N.
Existing optical systems (SAURON, Bacon et al., 2001; OASIS, McDermid et al., 2004)
have relatively low dispersion due to grism limitations, although the grisms allow for very
compact, undeviated systems. Grating-dispersed systems do exists in the NIR (OSIRIS,
Larkin et al., 2003). Future systems with VPH grisms and gratings will have even higher
efficiency; the coupling mode is well suited to articulated camera spectrographs. The
systems summarized in Table 3.4 are designed to exploit superb image quality with
fine spatial sampling (OASIS and OSIRIS are coupled to AO). While they cannot take
advantage of high dispersion without becoming read-noise limited, systems with larger
specific grasp could be optimized for high spectral resolution.
8 arcmin slit
3750 pixels
>λ
1’ slitlets
1.4’ spacing θ
>
>θ
δλ
<
7.5’’ slitlets
FMS 17’’ spacing
Figure 3.11. Filtered multi-slit schematic for SALT’s Robert Stobie Spectrograph (RSS). The
five panels on the left left show the progression from long slit, to filtered long slit, to two different
grids of filtered multi slits. Both are tuned to a high dispersion, 10 nm bandpass, and achieve
a spatial multiplex gain of 3 over a pure long slit, with a ×6 loss in bandpass. Higher spatial
multiplex (two to 10 times) is achieved at lower spectral resolution. The RSS Fabry–Perot mode
is shown for reference. The two rightmost panels show an overlay on a nearby, face-on galaxy
and some on-telescope calibration data for that slit-mask.
What is described here is more like the multi-object mode, but instead uses a regular
grid of slitlets. This is well-suited, for example, to observing single, extended sources.
An example of this type of instrument is the SALT Robert Stobie Spectrograph (RSS),
a prime focus imaging spectrograph with an 8 arcmin field of view, articulating camera,
VPH grating suite, dual Fabry–Perot etalons, and R = 50 order-blocking filters (Burgh
et al., 2003; Kobulnicky et al., 2003). The latter can be used with the multi-slit masks
to gain a factor of 3 in spatial multiplex at the highest spectral resolutions (R = 10,000
with a 10 nm bandpass). At lower resolutions (and fixed bandpass), the slit-packing can
be increased by factors of 2 to 10, such that the gain in spatial multiplex is comparable
to the loss of a factor of 5–6 in spectral multiplex in this particular case (the system is
designed for large spectral multiplex). Even at high spectral resolution what is gained
– beyond the spatial multiplex – is the ability to gain 2D spatial mapping in a single
exposure. On balance, what is lost and gained is comparable from a purely information
standpoint, and hence the choice is, as always, science driven. For the study of nearby
galaxy kinematics, this is an outstanding approach.
Sky subtraction
There are four primary issues concerning, and that are root causes of, sky subtraction
problems in spectroscopy: (i) low dispersion (sky lines contribute overwhelming shot-
noise); (ii) aberrations and non-locality (skyline profiles vary with field angle – spectral
and spatial – and time); (iii) stability (instrument); and (iv) under-sampling (compound
problems of field-dependent aberrations and flexure). All of these conditions are further
compounded if there is fringing on the CCD.
The solutions to these problems are instrumental, observational and algorithmic, i.e.
in the approach to the data analysis. The instrumental solution involves having a well-
sampled, high-resolution and stable system (you get what you pay for). Fibre-based
systems offer the most mapping flexibility, which is critical for spectrographs with
aberration-limited spectral image-quality. Pupil imaging (lenslets with or without fibres)
may offer advantages for HET/SALT-style telescopes, again if sky subtraction is spec-
trograph aberration-limited.
110 Matthew A. Bershady
The observational approach includes (i) beam-switching, where object and sky expo-
sures are interleaved, and (ii) nod-and-shuffle, where charge is shuffled on the detector in
concert with telescope nods between object and sky positions. Both of these approaches
have a 50% efficiency in either on-source exposure or in the number of sources that can
be observed (the on-detector source packing fraction).
An algorithmic approach entails aberration modelling, which is well suited to any of the
coupling methods that feed a spectrograph in a pseudo long slit. The question is to what
extent data analysis can compensate for instrumental limitations and avoid inefficient
observational protocol.
Some examples exist of telescope-time-efficient sky-subtraction algorithms; these are
solutions that do not require beam-switching or nod-and-shuffle. For example, Lissandrini
et al. (1994) identify flux and wavelength calibration, as well as scattered light, as the
dominant problems in their fibre-fed spectroscopic data. They use sky lines for second-
order flux calibration (after flat fields), model scattered light from neighbouring fibres
and map image distortions in pixel space to obtain accurate wavelength calibration. The
improvement is dramatic. Bershady et al. (2005) show that higher-order aberrations are
important; wavelength calibration is critical, but so too is line shape. They describe a
recipe for subtracting the continuum and fitting each spectral channel with a low-order
polynomial in the spatial dimension of the datacube. The algorithm works spectacularly
well for sources with narrow line emission with significant spectral channel offsets (e.g.
high internal dispersion as in a rapidly rotating galaxy, or intrinsically large velocity range
as in a redshift survey) and well-sampled data. For other instruments or sources (poor
sampling, low dispersion, broad lines, small velocity range), if aberrations are significant,
more dedicated sky fibres are needed. On balance, the optical stability of the instrument
is critical.
Are these post-facto algorithmic solutions 100% efficient? Not quite. One still needs
to sample sky, but, as derived in Bershady et al. (2004), the fraction of spatial elements
devoted to sky is relatively low (under 10%, and falling below 3% when the number of
spatial resolution elements exceeds 1000). So here is a case where, with a stable spectro-
graph, considerable efficiency may be gained by employing the right processing algorithm.
Consequently, fibre-fed, bench-mounted spectrographs offer the greatest opportunities to
realize these gains. Regardless of spectrograph type and feed, attention to modelling
optical aberrations is critical for good sky subtraction (Kelson, 2003; Viton and Milliard,
2003).
θ max θ max
∆λ
l n ∆λ θ
FP (or Michelson)
λ/λ0 ∼ θ 2
∆ path = 2 nl cos θ
Figure 3.12. Basic concept of etalon interference (a), and FP versus grating spectral resolution
as a function of angular aperture, θ (b).
dimension (bandpass or field sampling on the sky) has been gained via the temporal
domain, i.e. multiple observations. In this sense FPI has not yet been implemented for
truly 3D spectroscopy.
The basic principles of FPI, in the context of astronomical monochromators, can be
found in Geake et al. (1959), Vaughan (1967) and many other references. We summarize
the salient aspects to highlight here the field-widened capabilities (we are indebted to R.
Reynolds for the structure of the formal development). We discuss and give examples of
the two different FPI applications noted above and sketch how one might balance spatial
and spectral multiplex in future 3D instruments.
3.3.2 FP monochromators
FPs are conventionally thought of as being used with collimated beams (a mini-review
of such instruments from that era is presented by Bland and Tully, 1989). In this case,
there is the classic radial wavelength dependence in the image plane. At low spectral
resolution the bandpass can be made nearly constant over a large field of view (Jones
et al., 2002), as follows.
One way to characterize an etalon is by the size of its ‘bull’s eye’, or Jacquinot spot
(Jacquinot, 1954). The bull’s eye refers to the physical
' angle θ such that λ0 /|λ0 − λθ | <
R, and is given by θm ax = cos−1 (1 − 1/R) ≃ 2/R. This quantity is independent of
the telescope and is a property of the etalon. By coupling to a telescope, it is possible
to modify the angular scale (α) sampled on the sky by the bull’s eye. Since A × Ω is
conserved, α = θDe /DT , where De is the etalon diameter and DT the telescope diameter.
FPs can, however, be used in converging (or diverging) beams, even near a focus
(Bland-Hawthorn et al., 2001). Some examples include the optical FP on the 3.6 m CFHT,
when used with the AO Bonnette (AOB)8 and the future F2T2, a near-infrared double-
etalon system for FLAMINGOS-2 (8 m Gemini; Eikenberry et al., 2004a; Scott et al.,
2006). Image information is preserved by sampling the beam at a downstream focus, but
the spectral resolution is lowered (for a given finesse) at any spatial location because
each field angle on the sky is mapped into a range of physical angles through the etalon.
The degradation is not particularly severe for lower-finesse etalons or very slow beams.
The FLAMINGOS-2 multi-conjugate adapative optics (MCAO) focus for F2T2 is f /30,
and the AOB F-P beam is f /40. If the total angular field of view is much smaller than
the beam angle, or the focus is made telecentric, the bandpass is constant across field
angles on the sky, and the system forms a highly uniform tunable filter. The AOB optics
are not telecentric; this produces a radial degradation in the resolution.
3.3.3 FP spectrometers
Alternatively, the full spectral information can be extracted at the loss of the spatial
information by placing the etalons at or near a telecentric focus and sampling the pupil
in a collimated beam. The Wisconsin Hα Mapper (WHAM; Reynolds et al., 1998) is
the only astronomical example of this type of instrument. In this instance, the light is
collimated after it passes through the etalons, never refocused, and a detector is placed
at the pupil formed by the collimator. Field position on the detector contains spectral
information: each radius corresponds to a different wavelength. This is similar to the
monochromator application, except in this case each radial location on the detector has
a superposition from all spatial locations on the sky within the instrument entrance
aperture.
3.3.4 3D FP spectrophotometers
Grating-dispersed FPI
Arguably the most interesting FP monochromator mode is to eliminate the order-blocking
filters and grating-disperse the output beam to separate the orders on to the detector to
increase the spectral multiplex. See, for example, le Coarer et al.’s (1995) description of
PYTHEAS. Baldry and Bland-Hawthorn (2000) work out a particularly compelling case
8
For an earlier incarnation on this telescope, see www.cfht.hawaii.edu/Instruments/
Spectroscopy/Fabry-Perot/ and Joncas and Roy (1984).
3D spectroscopy instrumentation 113
for a cross-dispersed echelle system. The gain in spectral multiplex does not necessarily
cost spatial multiplex. In practice, some FPs are in spectrographs where they underfill
the detector and usable field in the image plane (e.g. RSS and F2T2). If the dispersion is
significantly greater than the etalon resolution, then, in addition to spectral multiplex,
this mode adds band-limited slitless spectroscopy in each FP order.
Pupil-imaging FPI
The above discussion frames the notion that detection downstream of an etalon at the
pupil of a collimated beam provides spectral information but no spatial information,
while detection at a focal surface provides the complement. A simple ray-trace shows
that between these two locations spectral and spatial information are mixed. By using
pupil imaging at the system input via a lenslet array (Section 3.2.9), detection at an
intermediate surface in a converging beam can separate spatial and spectral information.
Although this has never been done, in principle this could balance spatial and spectral
multiplex and allow for true 3D spectroscopy in future, field-widened instruments.
9
The presentation here benefited from discussion with J. Harlander, A. Sheinis, R. Reynolds,
F. Roesler and E. Merkowitz.
3D spectroscopy instrumentation 115
a) b)
Figure 3.13. Total and specific grasp versus spectral power for a range of instruments on 4-m-
and 10-m-class telescopes (solid and dashed lines, respectively) partially updated from Bershady
et al. (2005). See text for comments on instrument efficiency.
geometries would eliminate the need to go off-axis and probably allow for larger field.
Another approach is a Mach–Zender-style interferometer (Douglas et al., 1990). The
latter requires twice the detector real estate for the same number of spectral resolution
elements.
The primary advantage of an SHS is that it allows for very high spectral resolution for
a given solid angle relative to a conventional, grating-dispersed spectrograph. The SHS
is field-widened like an FP. This means the SHS can be built for low cost even on large
telescopes because the optics are small.
However, because the signal is in the form of an interferogram, there is what is known
as the ‘multiplex disadvantage.’ This can be expressed as the S/N performance of the
SHS relative to a grating spectrograph: S/NSHS = S/NGS (f /2)1/2 (SSHS /SGS )1/2 , where
S/NSHS and S/NGS are the signal to noise in SHS and grating spectrometer, respectively,
SSHS and SGS are the total photon signal, respectively, and f is the fraction of total signal
in a given spectral channel (f < 1, and decreases with bandwidth). In words, this means
that an SHS loses competitiveness with grating-dispersed spectrographs when the band-
pass is large. This has implications for design and use. Clearly, one must make SSHS and
f as large as possible. The small compact optics of an SHS system lend themselves to effi-
ciency optimization. To make f as large as possible, one must choose a small bandwidth
(but more than a Fabry–Perot monochromator) and remove OH lines via pre-filtering, or
by selecting bandpasses between them. Returning to Figure 3.1, SHS is between an FP
monochromator and other IFS methods and will therefore have application to a broad
range of science programmes that seek high spectral resolution over a limited bandpass
with good spatial coverage.
Figure 3.14. Spatial resolution (a) and specific grasp (b) versus spectral power for all instru-
ments in Tables 3.1–3.4, highlighting differences between optical (filled symbols) and NIR (open
symbols), as well as between different coupling methods (labelled).
Figure 3.15. Total grasp versus spectral power (a) and the number of spatial (NΩ ) versus
spectral (NR ) resolution elements for all instruments in Tables 3.1–3.4, highlighting differences
between optical (filled symbols) and NIR (open symbols). Dashed lines are at constant (labelled)
total information (NR × NΩ ).
Figure 3.14 shows that spatial resolution is higher in NIR instruments, while spectral
resolution is higher in optical instruments. Fibre IFUs have the largest specific grasp,
reflected in the bifurcation seen in spatial resolution, i.e. fibre-fed instruments have large
footprints per element (dΩ). There is a trend of decreasing specific grasp going from fibre
+lenslet, lenslet, and finally to slicers. ESI has unusually large A × dΩ for a slicer; RSS
in FMS mode has the highest specific grasp overall.
Figures 3.14 and 3.15 together show that optical and near-infrared instruments trade
spatial resolution for grasp; there are no high-grasp NIR instruments; the highest spectral
power instruments are optical. Optical and near-infrared instruments sample comparable
total information, with optical instruments sampling a broader range of trades between
spatial versus spectral information. Older NIR instruments clearly suffer from being
3D spectroscopy instrumentation 117
detector-size limited. IMACS-IFU stands out as having a significantly larger number of
total information elements, NR × NΩ , and in this sense is on par with future-generation
instruments.
10
See also the AURA MAXAT Final Report (1999), www.gemini.edu/science/maxat.
3D spectroscopy instrumentation 119
DT Ω dΩ
Instrument Coupling (m) (arcsec2 ) NΩ ∆λ/λ R NR ϵ
IRMOS slicer 30. 40. 0.01 4000 0.25 2000 500 ···
IRMOS slicer 30. 40. 0.01 4000 0.25 10000 2500 ···
IRIS slicer 30. 0.26 1.6e-5 16384 0.05 4000 200 ···
IRIS slicer 30. 1.33 8.1e-5 16384 0.05 4000 200 ···
IRIS slicer 30. 7.93 4.8e-4 16384 0.05 4000 200 ···
IRIS slicer 30. 41.0 2.4e-4 16384 0.05 4000 200 ···
WFOS fibre+lens 30. 810. 0.56 1440 1.37 5000 6850 0.3
11
See also www.tmt.org/observatory/index.html.
120 Matthew A. Bershady
DT Ω dΩ
Instrument Coupling Tel. (m) (arcsec2 ) NΩ ∆λ/λ R NR ϵ
and (iii) WFOS, an optical wide-field, seeing-limited spectrograph with potential for a
modest-grasp IFU with good spectral power and spectral resolution (R < 6000).
With the exception of WFOS, instrument design is driven by AO considerations
because of the enormous physical size of the image (which scales with mirror diameter for
a constant f -ratio). WFOS is necessarily a monster. The AO-driven focus is suitable for
scientific studies of un- or under-resolved sources (stars, planets, sub-kpc scales in distant
galaxies), and excellent science cases have been developed. Of this excellence there is no
doubt. Of concern is that once wedded to the notion of a very large telescope with no
clear path to building affordable, comparably monstrous instruments, one is forced down
a path, ab initio, of considering only science enabled by high angular resolution. It is not
surprising to note that WFOS – the one non-AO corrected instrument – stands out as
also the one TMT instrument concept that breaks into the high specific-grasp domain at
modest spectral-resolution domain (Figure 3.16). Indeed, as seen in Figure 3.17, WFOS
breaks new ground in terms of its total grasp with a higher spectral power than any
existing instrument (save ESI). To optimize low surface brightness studies, other paths
will need to be forged to push to higher spectral power and resolution at comparably
high grasp.
These same trends are also being played out for instrument design for extremely large
telescopes (ELTs) (e.g. Eisenhauer et al., 2000; Russell et al., 2004). We have focused
on TMT because of the more mature stage of this telescope’s planning. No doubt ELT’s
complement of instruments will open up exciting new capabilities, as demonstrated by the
superb forefront instrumentation on the VLT. The TMT instrumentation programme,
like that of the European ELT, is evolving rapidly. What is presented here is a snapshot
circa late 2005.
Figure 3.16. Spatial resolution (a) and specific grasp (b) versus spectral power including future
instruments in Tables 3.5 and 3.6: existing (filled circles), future ground-based 10 m-class tele-
scopes (open circles), future TMT (open squares), and future space-borne (open triangles).
The SNAP IFU (Ealet et al., 2003) is designed to identify SNe type out to z ∼ 1.7. As
such, it is unique in being a dual optical–NIR system (0.35–1.7 microns), with a 3 × 3
field using AIS, but very low spectral resolution (R = 100). With its very high expected
efficiency, co-added datasets should yield superb, spatially resolved spectrophotometry
of galaxies on 1–2 kpc scales.
Overall, future space-borne capabilities can be characterized as having 3 × 3 arcsec2
fields mapped with AIS technology with 0.15 arcsec sampling, lower spatial resolution
than TMT. Spectral resolution is in the 100 < R < 3000 range, again lower than TMT.
This is consistent with their being competitive in performance relative to TMT-class
instruments, given Gillett and Mountain’s (1998) argument. However, there are no large-
grasp systems that take full advantage of the low backgrounds of space. There are no
high- or even medium-resolution spectrographs to couple, competitively, to such large
angular apertures. Nonetheless, barring past fiascos, the space-borne missions offer the
guarantee of superlative image quality and low backgrounds extending into the mid-IR,
while ground-based observatories face the intense challenge of developing advanced AO
systems.
Figure 3.17. Total grasp versus spectral power (a) and the number of spatial (NΩ ) versus
spectral (NR ) resolution element (b) including future instruments in Tables 3.5 and 3.6; symbols
as in previous figure.
Acknowledgements
We would like to thank the Instituto de Astrofı́sica de Canarias (IAC) and Winter School
organizers, the University of Toronto Department of Astronomy and Astrophysics for
their gracious hospitality during a sabbatical year where this work was done, and the
NSF for their financial support of this research (AST-0307417 and AST-0607516).
REFERENCES
Afanasiev V.L., Moiseev, A.V. (2005), Astr. Lett., 31, 193
Afanasiev, V.L, Dodonov, S.N., Sil’chenko, O.K., Vlasyuk, V.V. (1990), SAO preprint,
N54
Aldering, G., Akerlof, C.W., Amanullah, R. ET AL. (2002), Proc. SPIE, 4835, 146
Alighieri, S.S. (2005), in the Proceedings of the ESO Workshop on Science Perspectives for
3D Spectroscopy, (Garching, Germany)
Allende Prieto, C., Majewski, S.R., Schiavon, R. ET AL. (2008) AN, 329, 1081
Allington-Smith, J., Content, R. (1998), PASP, 110, 1216
Allington-Smith, J., Murray, G., Content, R. ET AL. (2002), PASP, 114, 892
Allington-Smith, J., Dubbeldam, C.M., Content, R. ET AL. (2004), Proc. SPIE, 5492,
701
Arribas, S., Mediavilla, E., Rasilla, J.L. (1991), ApJ, 369, 260
Arribas, S., del Burgo, C., Carter, D. ET AL. (1998), ASPCS, 152, 149
Atherton, P.D., Taylor, K., Pike, C.D., Harmer, C.F.W., Parker, N.M., Hook, R.N.
(1982), 201, 661
Avila, G., Guinouard, I., Jocou, L., Guillon, F., Balsamo, F. (2003), Proc. SPIE, 4841,
997
Bacon, R., Copin, Y., Monnet, G. ET AL. (2001), MNRAS, 326, 23
Bacon, R., Bauer, S.-M., Bower, R. ET AL. (2004), Proc. SPIE, 5492, 1145
3D spectroscopy instrumentation 123
Baldry, I.K., Bland-Hawthorn, J. (2000), PASP, 112, 1112
Baldry, I.K., Bland-Hawthorn, J., Robertson, J.G. (2004), PASP, 116, 403
Baranne, A. (1972), in ESO/CERN Conference on Auxiliary Instrumentation for Large Tele-
scopes, ed. S. Lautse & A. Reiz (Geneva), 227
Barden, S.C., Wade (1988), ASPCS, 3, 113
Barden, S.C., Elston, R., Armandroff, T., Pryor, C. (1993), ASPCS, 37, 223
Barden, S.C., Sawyer, D.G., Honneycutt, R.K. (1998), Proc. SPIE, 3355, 892
Barden, S.C., Arns, J.A., Colburn, W.S., Williams, J.B. (2000), PASP, 112, 809
Bershady, M.A., Andersen, D.R., Harker, J., Ramsey, L.W., Verheijenn, M.A.W.
(2004), PASP, 116, 565
Bershady, M.A., Andersen, D.R., Verheijen, M.A.W., Westfall, K.B., Crawford,
S.M., Swaters, R.A. (2005), ApJS, 156, 311
Bershady, M., Barden, S., Blancho, P.-A. ET AL. (2008) Proc. SPIE, 7014, 70140H-1
Blais-Ouellette, S., Guzman, D., Elgamil, A., Rallison, R. (2004), Proc. SPIE, 5494,
278
Blais-Ouellette, S., Daigle, O., Taylor, K. (2006), Proc. SPIE, 6269, 174
Bland, J., Tully, R.B. (1989), AJ, 98, 723
Bland-Hawthorn, J., van Breugel, W., Gillingham, P.R., Baldry, I.K. (2001), ApJ,
563, 611
Bonnet, H. ET AL. (2004), Msngr, 117, 17
Boulesteix, J., Georgelin, Y., Marcelin, M., Monnet, G. (1984), Proc. SPIE, 445, 37
Burgh, E.B., Nordsieck, K.H., Kobulnicky, H.A., Williams, T.B., O’Donoghue, D.,
Smith, M.P., Percival, J.W. (2003), Proc. SPIE, 4841, 1463
Burgh, E.B., Bershady, M.A., Westfall, K.B., Nordsieck, K.H. (2007), PASP, 119, 1069
Carrasco, E., Parry, I.R. (1994), MNRAS, 271, 1
Clenet, Y., Le Coarer, E., Joncas, G. ET AL. (2002), PASP, 114, 563
Crampton, D., Simard, L. (2006), Proc. SPIE, 6269, 59
Davila, P.S., Bos, B.J., Contreras, J. ET AL. (2004), Proc. SPIE, 5487, 611
Dopita, M.A., Waldron, L.E., McGregor, P. ET AL. (2004), Proc. SPIE, 5492, 262
Douglas, N.G., Butcher, H.R., Melis, M.A. (1990), Ap&SS, 171, 307
Dressler, A., Hare, T., Bigelow, B.C., Osip, D.J. (2006), Proc. SPIE, 6269, 13
Ealet, A., Prieto, E., Bonissent, A. ET AL. (2003), Proc. SPIE, 4850, 1169
Eikenberry, S.S., Elston, R., Raines, S.N. ET AL. (2004a), Proc. SPIE, 5492, 1196
Eikenberry, S.S., Elston, R., Guzman, R. ET AL. (2004b), Proc. SPIE, 5492, 1264
Eisenhauer, F., Tecza, M., Thatte, N., Mengel, S., Hofmann, R., Genzel, R. (2000),
Proceedings of the Backaskog Workshop on Extremely Large Telescopes, ed. T. Andersen,
A. Ardeberg, R. Gilmozzi (Lund/ESO), 57, 292
Eisenhauer, F., Abuter, R., Bickert, K. ET AL. (2003), Proc. SPIE, 4841, 1548
Fabricant, D.G., Hertz, E.N., Szentgyorgyi, A.H. ET AL. (1998), Proc. SPIE, 3355, 285
Fabricant, D.G., Fata, R., Roll, J. ET AL. (2005), PASP, 117, 1411
Flores, H., Peuch, M., Hammer, F., Garrido, O., Hernandez, O. (2004), A&A, 420,
L31
Geake, J.E., Ring, J., Woolf, N.J. (1959), MNRAS, 119, 161
Gelderman, R., Woodgate, B.E., Brown, L.W. (1995), ASPC, 71, 89
Gillett, F., Mountain, M. (1998), ASPC, 133, 42
Gordon, S., Koribalski, B., Houghton, S., Jones, K. (2000), MNRAS, 315, 248
Guerin, J., Felenbok, P. (1988), ASPCS, 3, 52
Hanuschik, R.W. (2003), A&A, 407, 1157
Harlander, J., Reynolds, R.J., Roseler, F.L. (1992), ApJ, 730
Hartung, M., Lidman, C., Ageorges, N., Marco, O., Kasper, M., Clenet, Y. (2004),
Proc. SPIE, 5492, 1531
Haynes, R., Lee, D., Allington-Smith, J. ET AL. (1999), PASP, 111, 1451
Hearty, F.R., Morse, J., Beland, S. ET AL. (2004), Proc. SPIE, 5492, 1623
Henault, F.R., Bacon, R., Content, R. ET AL. (2004), Proc. SPIE, 5249, 134
Herbst, T.M., Beckwidth, S. (1988), PASP, 100, 635
Hernandez, O., Gach, J.-L., Carignan, C., Boulesteix, J. (2003), Proc. SPIE, 4841,
1472
Hill, G.J., MacQueen, P.J., Tejada, C. ET AL. (2004), Proc. SPIE, 5492, 251
Iserlohe, C., Tecza, M., Eisenhauer, F. ET AL. (2004), Proc. SPIE, 5492, 1123
124 Matthew A. Bershady
Jacquinot, P. (1954), JOSAA, 44, 761
Joncas, G., Roy, J.R. (1984), PASP, 96, 263
Jones, D.H., Shopbell, P.L., Bland-Hawthorn, J. (2002), MNRAS, 329, 759
Kelson, D. (2003), PASP, 115, 688
Kelz, A., Verheijen, M.A.W., Roth, M.M. ET AL. (2006), PASP, 118, 119
Kenworthy, M.A., Parry, I.R., Ennico, K.A. (1998), ASPCS, 152, 300
Kenworthy, M.A., Parry, I.R., Taylor, K. (2001), PASP, 113, 215
Kobulnicky, H.A., Norsieck, K.H., Burgh, E.B., Smith, M.P., Percival, J.W.,
Williams, T.B., O’Donoghue, D. (2003), Proc. SPIE, 4841, 1634
Koo, D.C., Bershady, M.A., Wirth, G.D., Stanford, S.A., Majewski, S.R. (1994), ApJ,
427, L9
Larkin, J.E., Quirrenbach, A., Krabbe, A. ET AL. (2003), Proc. SPIE, 4841, 1600
Le Coarer, E., Bensammar, S., Comte, G., Gach, J.L., Georgelin, Y. (1995), A&AS,
111, 359
Le Fevre, O., Crampton, D., Felenbok, P., Monnet, G. (1994), A&A, 282, 325
Le Fevre, O., Saisse, M., Mancini, D. ET AL. (2003), Proc. SPIE, 4841, 1671
Lissandrini, C., Cristiani, S., La Franca, F. (1994), PASP, 106, 1157
Lorenzetti, D., Cortecchia, F., D’Alessio, F. ET AL. (2003), Proc. SPIE, 4841, 94
Maihara, T., Iwamuro, F., Yamashita, T. ET AL. (1993), PASP, 105, 940
McDermid, R., Bacon, R., Adam, G., Benn, C., Cappellari, M. (2004), Proc. SPIE, 5492,
822
McGregor, P.J., Hart, J., Conroy, P.G. ET AL. (2003), Proc. SPIE, 4841, 1581
Murphy, T.W., Matthews, K., Soifer, B.T. (1999), PASP, 111, 1176
Nelson, G.W. (1988), ASPCS, 3, 2
Oliveira, A.C., de Oliveira, L.S., dos Santos, J.B. (2005), MNRAS, 356, 1079
Parry, I., Bunker, A., Dean, A. ET AL. (2004), Proc. SPIE, 5492, 1135
Prieto, E., Ferruit, P., Cuby, J.-G., Blanc, P.-E., Le Fevre, O. (2004), Proc. SPIE,
5487, 777
Rallison, R.D., Schicker, S.R. (1992), Proc. SPIE, 1667, 266
Ramsay Howat, S.K., Todd, S., Wells, M., Hastings, P. (2006), New Astronomy Reviews,
50, 3513
Ramsey, L.W. (1988), ASPCS, 3, 26
Ren, D. & Allington-Smith, J. (2002), PASP, 114, 866
Reynolds, R.J., Tufte, S.L., Haffner, L.M., Jaehnig, K., Percival, J.W. (1998), PASA,
15, 14
Roche, P.F., Lucas, P.W., Mackay, C.D. ET AL. (2003), SPIE, 4841, 901
Roesler, F.L. (1974), Methods in Experimental Physics, 12A, Academic Press (San Diego),
Chapter 12
Rosado, M., Langarica, R., Bernal, A. ET AL. (1995), RMxAC, 3, 263
Rosado, M., Cruz-Gonzales, I., Salas, L. ET AL. (1998), SPIE, 3354, 1111
Roth, M.M., Kelz, A., Fechner, T. ET AL. (2005), PASP, 117, 620
Russell, A.P.G., Monnet, G., Quirrenbach, A. ET AL. (2004), Proc. SPIE, 5492, 1796
Saunders, W., Bridges, T., Gillingham, P. ET AL. (2004), Proc. SPIE, 5492, 389
Schmoll, J., Roth, M.M., Laux, U. (2003), PASP, 115, 854
Schmoll, J., Dodsworth, G.N., Content, R., Allington-Smith, J.R. (2004), Proc. SPIE,
5492, 624
Schroeder, D. (2000), Astronomical Optics, Academic Press (San Diego), 2nd Edition
Scott, A., Javed, M., Abraham, R. ET AL. (2006), Proc. SPIE, 6269, 176
Sharples, R.M., Bender, R., Lehnert, M.D. ET AL. (2004), Proc. SPIE, 5492, 1179
Sheinis, A.I., Bolte, M., Epps, H.W. ET AL. (2002), PASP, 114, 851
Sheinis, A.I. ET AL. (2006), Proc. SPIE, (astro-ph/0606176)
Sluit, A.P.N., Williams, T.B. (2006), AJ, 131, 2089
Tamura, N., Murray, G.J., Sharples, R.M., Robertson, D.J., Allington-Smith, J.R.
(2005), OExpr, 13, 4125 (astro-ph/0509913)
Tanaka, M., Yamashita, T., Sato, S., Okuda, H. (1985), PASP, 97, 1020
Thatte, N.A., Weitzel, L., Cameron, M. ET AL. (1994), Proc. SPIE, 2224, 279
Thatte, N., Anders, S., Eisenhauer, F. ET AL. (2000), ASPCS, 195, 206
Thorne, A.P. (1988), Spectrophysics, Chapman & Hall (London)
Tull, R.G., MacQueen, P.J., Sneden, C., Lambert, D.L. (1995), PASP, 107, 251
3D spectroscopy instrumentation 125
Vaughan, A.H. (1967), ARA&A, 5, 139
Verheijen, M.A.W., Bershady, M.A., Andersen, D.R. ET AL. (2004), AN, 325, 151
Viton, M., Milliard, B. (2003), PASP, 115, 243
Williams, T.B., Nordsieck, K.H., Reynolds, R.J., Burgh, E.B. (2002), ASPC, 282,
441
Wright, G.S., Rieke, G.H., Colina, L. ET AL. (2004), SPIE, 5487, 653
Wynne, C.G. (1991), MNRAS, 250, 796
Wynne, C.G., Worswick, S.P. (1989), MNRAS, 237, 239
Yee, H.K.C., Ellingson, E., Carlberg, R.G. (1996), ApJS, 102, 269
4. Analysis of 3D data
PIERRE FERRUIT
4.1 Introduction
4.1.1 Presentation and scope
This chapter contains the proceedings of the course on analysis of 3D spectrographic
data given as part of the XVII Canary Island Winter School of Astrophysics. It provides
an overview of some basic and generic analysis techniques for 3D spectrographic data.
It includes a description of an arbitrary selection of tasks with, whenever possible,
examples on real data and a lot of discussion about noise and errors. To illustrate the
examples, we will make a heavy use of tools that are part of the XOasis software developed
at the Centre de Recherche Astrophysique de Lyon (CRAL) and of 3D datasets obtained
using the TIGER and OASIS instruments.
This course is not limited to pure 3D analysis techniques as the core of the analysis of
3D datasets is either identical or similar to what is done for regular spectrographic data.
! It is not exhaustive but contains a rather arbitrary selection of tasks and tools that
It has some obvious caveats and limitations:
! It is limited to the data analysis techniques used in visible and near-infrared (NIR)
! It does not address those used in the radio and X-ray communities (long-time users
astronomy.
of 3D spectrography).
Figure 4.2. Sketch showing a (non-exhaustive) set of tasks, and an unsuccessful attempt to
sort them between data processing and analysis tasks.
128 Pierre Ferruit
associated noises and how they propagate through the various steps of analysis. We will
use very simple examples to illustrate how this can be done and in the following we will
set the stage and give the basic definitions of the quantities that will be used throughout
this document.
4.2.2 Noises
! the instrument, e.g. through its detector. In the case of the visible and near-infrared
We will distinguish between four different sources of noise in a dataset:
detectors on which our examples will be based, typical detector noises are the read-
! undesired backgrounds like sky emission. Their shot noise will remain even if they
out noise and the shot noise from the detector dark current.
! the signal itself. The signal will have its own shot noise.
have been perfectly subtracted.
! the data reduction. Noise is also introduced in the data during its reduction/
processing, due to the finite signal to noise ratio of the calibration exposures.
$ N
1
Vx = ⟨(x − x)2 ⟩ = × (xi − x)2 (4.2)
N i=1
with
$ N
1
COVy z = ⟨(y − y) × (z − z)⟩ = × (yi − y) × (zi − z) (4.10)
N i=1
If the statistical variables are uncorrelated, then the covariance will be zero.
$ N $ N
1
F = × Fi × δλi with ∆λ = δλi (4.11)
∆λ i=1 i=1
Using the relations listed in the previous section, we can derive the average of the sta-
tistical variable F as a function of those of the Fi :
$ N $ N
1
F = × F i × δλi with ∆λ = δλi (4.12)
∆λ i=1 i=1
Here we have assumed that the size of the spectral pixels was perfectly known. If, in
addition, we assume that the Fi are uncorrelated (i.e. their covariance is zero), we have:
N
$ ) *2
δλi
V = Vi × (4.13)
i=1
∆λ
) * 1
,
S F × N F i × δλi
= √ = ∆.λ, i=1 / 0 (4.14)
N V N 2
V × δλi
i=1 i ∆λ
We can immediately see that the contribution of each spectral pixel to the final variance
is weighted by the square of the size of the spectral pixel.
130 Pierre Ferruit
If we now assume that we have a constant spectral pixel size (δλi = cte = δλ; ∆λ =
N × δλ) and that we are limited by the photon noise of the object (Vi = F i ), this yields
the following simplified relations:
1 $
N N
1 $ S √ '
F = F i, V = 2 F i and = N× F (4.15)
N i=1 N i=1 N
We find the classical Poissonian statistics gain in signal to noise ratio proportional to the
square root of N and this example shows how the signal to noise can be tracked when
performing an operation on the data.
where F (λ) is the final averaged spectrum and γ(λ) is the spectrophotometric correction
factor that is applied to all the uncalibrated spectra of each exposure Γi (λ).
The relation giving the variance of F (λ) as a function of the variance of γ(λ) and of
the variance of the uncalibrated data is as follows:
N ) *2 (N +2
$ γ(λ) Vγ (λ) $
V (λ) = Vi (λ) × + × Γi (λ) (4.17)
i=1
N N2 i=1
This shows that the spectrophotometric calibration of the data has introduced a system-
atic error term that does not decrease when the datasets are averaged. This means that
after a certain limit, increasing the number of exposures N will not help to increase the
signal to noise ratio. This latter becomes limited by the constant term in the variance.
This constant, systematic error term is due to the fact that the same spectrophoto-
metric correction factor has been used for the spectra of all exposures, introducing a
correlated source of noise in the data. A common way to solve this problem is to place
the object at different locations in each exposure. In this case, the correction factor is
now exposure dependent and becomes γi (λ). We will however assume that its variance
is uniform over the instrument field of view, i.e. Vγ i (λ) = Vγ (λ). The initial relation now
becomes:
$ N $ N
1 1
F (λ) = × Fi (λ) = × γi (λ) × Γi (λ) (4.19)
N i=1
N i=1
The constant term in the variance has now disappeared and has been replaced by a term
decreasing as 1/N . By changing the observation strategy, it has been possible to avoid
introducing a systematic error term. This illustrates how keeping track of the noise can
help in devising better observation, processing or analysis strategies.
132 Pierre Ferruit
Figure 4.3. Sketch illustrating how an artifact can be spread over neighbouring spaxels when
resampling spatially a dataset.
4.2.4 Artifacts
Artifacts are one last item in our list of features that will affect data. Typical sources of
! the cosmic ray impacts on the detector and the residuals of their removal; and
! the residuals from the sky subtraction (see Section 4.4).
As for the noise, it is important to keep track of pixels affected by artifacts. In the Euro3D
format, this is done through the use of a data quality flag for each spectral pixel of each
spectrum of the datacube (see Section 4 of Kissler-Patig et al., 2003).
! Many 3D spectrographs have sampling patterns that are not a regular grid (e.g. in
The spatial sampling provided by the spectrograph can be irregular in different ways:
! There can be holes in the field of view coverage even when the sampling is not
intrinsically sparse (observation that is the sum of non-contiguous fields of view,
rejected spaxels, etc.).
Analysis of 3D data 133
Figure 4.4. Images showing the location of the spaxels within the field of view of three different
3D spectrographs (SAURON, OASIS and PMAS in its PPAK mode).
Figure 4.5. Example from the image reconstruction and visualization of a datacube with-
out (left panel) and with (right panel) spatial resampling. The figure has been extracted from
Sánchez (2004) and corresponds to a dataset of the INTEGRAL 3D spectrograph, which has
circular, non-contiguous spaxels.
Examples of the spatial distribution of spaxels for various 3D spectrographs are shown
in Figure 4.4.
When reconstructing an image from a datacube, two different approaches can be used
depending on the adopted philosophy about the resampling of data. To avoid resampling,
a special visualization tool like the E3D software (see tutorial on the use of E3D in this
book) can be used that will take into account the location and shape of the spaxels and
display them without resampling (see left panel of Figure 4.5).
134 Pierre Ferruit
Figure 4.6. Example of image reconstruction procedure using the XOasis software.
The right panel of Figure 4.5 shows the results obtained for the same dataset when
resampling and interpolation of the data are used (the black circles superimposed on the
image are the shape and position of the spaxels). In Figure 4.6, we show screenshots of
the graphical user interface (GUI) associated to the image reconstruction tools of the
Analysis of 3D data 135
XOasis software suite. It clearly shows the steps to be followed from a datacube to an
! The first step is the ‘collapsing’ of the spectra along the wavelength direction (upper
image:
panel of Figure 4.6). The user defines a wavelength range (that can be the whole
spectrum) over which the spectral pixels will be averaged or summed (possibly using
weighted summation). As an output, a table is produced that – for each spaxel of
the datacube – contains its position and the value obtained when collapsing the
! The second step (central panel of Figure 4.6) is the actual image reconstruction
corresponding spectrum.
that will produce an image with square pixels and a regular sampling from the table
produced during the first step. A triangulation will first be performed (looking for the
nearest neighbours of each new sampling point in the table) and will be followed by
the interpolation itself. The user is provided with the choice of allowing extrapolation
of data and/or filling in the large gaps between spaxels by interpolation. In the
example we have chosen extrapolation was not allowed, and the output image values
outside the original field of view are filled with zero values (see bottom panel of
Figure 4.6).
In much software (e.g. E3D) the user is also provided with a choice of interpolation
methods.
Figure 4.7. Snapshot of datacube visualization using the VolView tools extracted from the
work of J. Gerssen.
20
18
16
Flux density (phot/s/nm/arcsec 2/m 2)
14
12
10
0 0 1 2 3 4 5 6
350 450 550 650 750 850 950 1050 1150
Wavelength (nm)
Figure 4.8. Spectra of the sky emission in the visible (left panel, from the GEMINI observatory)
and of the zodiacal light (right panel).
Obviously, all the atmospheric background will disappear in the case of observations
in space. The dominant external background then becomes the (much fainter) zodiacal
light originating from the dust localized in the ecliptic plane of the Solar System. It is
a mix of scattering of solar light (toward the blue) and thermal emission from the dust
(toward the red). Typical sky and zodiacal light spectra are shown in Figure 4.8.
T = S + B + d and Ti = Bi + di (4.21)
where S is the signal, B and Bi the background and d and di the detector dark current.
The associated variances are (assuming a ‘shot-noise’ type of contributions for the signal,
background and dark current):
i i
VT = S + B + d + σ 2 and VTi = B + d + σ 2 (4.22)
! The instrument spectral point-spread function (PSF) can differ from one spaxel to
the case, can be a problem. This is mainly due to two different effects:
the other. As an example, the spectral PSF of an integral field unit (IFU) quite often
! The wavelength calibration of a datacube is not perfect and there may be shifts
varies within its field of view.
0 −0.001
0.1
−0.002
0 −0.002
−4 −2 0 2 4 −4 −2 0 2 4 −4 −2 0 2 4
Position ( ) Position ( ) Position ( )
Figure 4.9. Left panel: plot of the nominal Gaussian line profile with the plots of the narrow/
broader and blueshifted/redshifted profiles superimposed. Central panel: residuals from the sub-
traction of the narrower/broader profiles from the nominal one. Right panel: residual from the
subtraction of the blueshifted/redshifted profiles from the nominal one.
Figure 4.10. Left panel: spectrum of an active galaxy in the 630–700 nm range displaying the
prominent emission lines from the object and an atmospheric absorption feature between 680
and 690 nm. Right panel: zoom on the atmospheric absorption feature (which is made of several
lines).
In the case of the subtraction of a background with a continuum spectrum, these effects
will have little impact. They can however be a source of artifacts when subtracting a
background with spectrally unresolved emission lines (i.e. the sky).
To illustrate this effect, we have performed a simple simulation. We have produced
an artificial emission line with a Gaussian profile (centred on zero and with a sigma
of unity; see Figure 4.9). We have then introduced very small changes in the Gaussian
profile parameters (changes in sigma of 1%; shifts of the centroid of the line by 1% of
the sigma of the Gaussian profile) and created the corresponding spectra. A plot with all
three profiles overlapped is displayed in Figure 4.9, and it shows that the lines are barely
distinguishable.
However, if we now subtract the lines with the modified Gaussian profile to the original
one, we see that we end up with negative and positive residuals that can reach 0.5–1.0%
of the peak intensity of the line (see Figure 4.9). Although the amplitude of the residuals
seems small, this can be a major issue when the background is stronger than the object,
as it is quite often the case for deep exposures and in the near-infrared.
Figure 4.11. Sketch illustrating how the spectrum of a point source can be distorted due to
the wavelength-dependent shift induced by atmospheric refraction.
One common way to determine the correction for this absorption is to calibrate it
using a featureless spectrum such as the spectrum of the spectrophotometric standard
star used for the radiometric calibration of the instrument (this latter being quite often
! Assume the stellar continuum is flat at the location of the absorption feature.
selected to have a spectrum as featureless as possible). The steps are as follows:
! Fit the absorption line and determine the relative absorption (with respect to the
! Apply the corresponding multiplicative correction to all the spectra of the datacube.
continuum level).
A major caveat is the fact that the importance of this absorption feature depends strongly
on the parameters of the observation (in particular the time and the zenithal angle), and
they can be quite different for the spectrophotometric star and the object itself. It is
therefore not usually possible to apply the correction directly, and one needs to scale it
to the object airmass and hope that time variation of the absorption between the two
observations has remained small.
Atmospheric refraction
One last atmospheric effect is atmospheric refraction. It will cause the image of an object
to be shifted as a function of wavelength. The effect decreases as wavelength increases
and it is usually negligible starting in the near-infrared.
In slit spectroscopy it will cause the object to gradually move outside of the slit as
wavelength changes. In 3D spectroscopy, there is no longer a risk of ‘losing’ the object (as
long as the field of view is large enough) as it will just ‘move’ within the field of view as a
function of wavelength. Reconstructing monochromatic images at different wavelengths
and comparing them will easily show the shift. However, if it is not corrected, this shift
will distort the spectra of the object, changing its slope from one spaxel to the other, as
illustrated in Figure 4.11.
A detailed description of the impact of atmospheric refraction and how to correct for
it can be found in Arribas et al. (1999). Typically, the steps of the correction procedure
are the following (and they are only possible when using a 3D spectrograph):
140 Pierre Ferruit
! Divide the datacube in monochromatic ‘slices’ (one image of the field of view per
! Resample all the slices spatially using a common sampling grid (see comments on
One difficulty can be the determination of the correct shift. It is possible to use theoretical
values or to use a point source within the field of view (if available), paying great attention
not to measure shifts intrinsic to the source itself.
Note that some instruments are now equipped with atmospheric refraction correctors
(ADC). This is almost mandatory when using 3D + adaptive optics in the visible.
! Obtaining a collection of observations of the same field of view with only a small
goals:
pointing offset between each exposure in order to improve the spatial sampling of
4.5.1 Drizzling
The drizzling technique has been advertised a lot within the astronomical community
through its use for HST (Hubble Space Telescope) images. It aims at improving the
sampling of the image and can be very useful for a diffraction-limited instrument oper-
ating over a large wavelength range. In this case, the size of the PSF will change quickly
with wavelength, and quite often the instruments cannot sample it correctly at all wave-
lengths and the PSF ends up undersampled in the blue; see Hook and Fruchter (2000)
for a description of the technique and its application to images.
So far it has only been applied to images but it can be naturally extended to 3D
datacubes (after all, they can be considered as stacks of monochromatic slices). However,
we are not aware of any drizzling software adapted to datacubes.
Figure 4.12. Reconstructed images for the 7 individual NGC 4258 datacubes (to the left)
obtained with the OASIS/CFHT instrument and for the final datacube (to the right) obtained
by co-adding (merging) them.
We will discuss some of these points in the following subsections and we will systemat-
ically use some OASIS/CFHT (Optically Adaptive System for Imaging Spectroscopy /
Canada–France–Hawaii Telescope) observations of the galaxy NGC 4258 as an example
(see Figure 4.12).
Relative positioning
Before trying to co-add the datacubes, it is necessary to determine their relative posi-
tioning. The easy way is of course to know the offset beforehand. This is rarely possible,
especially as many observations have sub-arcsecond samplings and it is necessary to reach
accuracies of a fraction of a spaxel size. Quite often, this information must therefore be
obtained from the data itself.
If a ‘sharp’ morphological feature (e.g. the nucleus of a galaxy, a star, etc.) is available
within the field of view it is possible to use the position of this source (as derived,
for example, from centroiding or Gaussian profile fitting) as a reference. If no ‘sharp’
morphological feature is available, it is also possible to use contour plots of reconstructed
images to get the relative positioning of two datacubes. This is illustrated in Figure 4.13.
! As it is, with inspection of the overlap being done by eye, this is a very subjective
The difficulties are usually as follows:
! Changes in observing conditions (e.g. a seeing change) will mess up with the contour
method.
Relative normalization
After registering the datacubes with respect to each other, we also have to check their
relative normalization, i.e. tracking differences in radiometric response of the complete
system (including the atmosphere if relevant) from one observation to another. In most
cases, a dedicated spectrophotometric calibration will not be available for each exposure
142 Pierre Ferruit
4 4
2 2
Position
Position
0 0
−2 −2
−4 −4
−6 −6
−5 0 5 −5 0 5
Position Position
Figure 4.13. Example of the use of contour plots of a reconstructed image to derive the relative
positioning of two datacubes. Left panel: superimposed contours of the two images for the initial
positioning. Right panel: same but for the final positioning.
and it is very difficult to track changes like a modification of the transparency of the
atmosphere. In the case of the deep-field approach, getting the relative normalization
between the datacubes is straightforward as they all cover the same field of view and
their flux or surface brightness levels can easily be compared.
Ready to co-add?
At this stage, we should have N datacubes that have been registered and normalized
with respect to each other and we should be ready to co-add (merge) them. However,
! Is it optimal in terms of signal to noise (which is the driver for the deep field
approach)?
In order to answer these questions, we are going to use a very simple and quite arti-
ficial example. It will show that performing a weighted summation when co-adding the
datacubes can help a lot. This allows us to take into account the fact that the datacubes
can show up differences in signal to noise ratio and reflects the quite obvious fact that it
is not desirable to give the same weight to low and high signal to noise datacubes.
We will assume that the calibration of the datacubes has introduced a division by the
integration times (to get the spectra in unit of power, i.e. energy per unit of time) and
that we are in photon noise regime. This yields:
) *i √
1 i S
i
Scalibrated = i × S i and V i = S and = Si (4.28)
t N
The variance of a calibrated datacube is:
) *2 ) *2
i
∂Vcalibrated i 1 1 i
i
Vcalibrated =V ×
i
=S × i = i × S calibrated (4.29)
∂V i t t
$ N N
1 S calibrated $ 1
Vcalibrated = 2
× i
Vcalibrated = × (4.31)
N i=1
N2 i=1
ti
) * .
S S calibrated 1
=√ = ., ×N × S calibrated (4.32)
N calibrated Vcalibrated N 1
i=1 t i
1 23 4
source of trouble
The sum of the inverse of the integration times that appear in the equation for the signal
to noise ratio is going to be a source of trouble. To see that, we are going to compare
two practical cases.
First, if all the integration times are the same (ti = t = cte):
) * . .
S N √ √
= . × S calibrated = t × N × S calibrated (4.33)
N calibrated N 1 23 4
t usual factor
Second, if we have two exposures with different integration times (ti=1 = t and ti=2 =
t/2):
) * . .
S 2 √ 2
= . × S calibrated = t × √ × S calibrated (4.34)
N calibrated 3 3
t 1 23 4
smaller factor
We can see that for exposures of different integration times, we do not have the expected
gain in signal to noise ratio (expected to be proportional to the square root of the
cumulated integration time). This is due to the contribution of the short exposure time
exposures. The solution is to put more weight on long exposures than on short ones.
With ttot being the sum of all individual integration times (cumulated integration time)
and using a normalized weight ti /ttot , we have:
N
$ N
$
weighted 1 1
Scalibrated = × i
t × i
Scalibrated = × Si (4.35)
ttot i=1
ttot i=1
N
$ N
$
weighted 1 1
Vcalibrated = × Vi = × Si (4.36)
t2tot i=1
t2tot i=1
) *weighted weighted .
S S √
= . calibrated = ttot × S calibrated (4.37)
N weighted
calibrated Vcalibrated
We have obtained the classical factor of square root of the cumulated integration time.
This simple and naive example outlines the fact that it is necessary to assess the impact
of each processing/analysis step on the signal to noise of the data. In this case the simple
fact of having calibrated data has forced us to modify the way the datacubes are co-added
(if data had still been in counts, a direct average would have worked). This explains as
an example why the HST images are never in physical units (per unit of time) by default.
It is simply to allow the user to perform direct summations on them without having to
worry about weights.
144 Pierre Ferruit
Figure 4.14. Example of graphical user interface (from the XOasis software) used as a front-
end for a datacube merging tool. It can be seen that this program allows use of a weighted
summation (with the weights set by default to ti /ttot ).
Figure 4.15. Spectra extracted from an individual datacube (left panel) and from the final
merged datacube (right panel) at the same location in the object.
The (very simple) procedure we have described is in fact quite similar (although less
sophisticated) than the ones used in optimal summation techniques.
! coping with changes of image quality and atmospheric transmission from one expo-
sure to another. Note that how to deal with changes in image quality is an entire
! coping with the intrinsic stability of the telescope plus instrument system. This is
subject in itself and is not addressed here;
critical for this type of observation, which tends to push the instruments to their
limits.
This type of prototype study is paving the way for future 3D spectrographic deep fields
that will be at the heart of the scientific goals of the future generation of 3D spectrographs
like MUSE (Multi-Unit Spectroscopic Explorer) and KMOS (K-Band Multi-Object Spec-
trometer) for the VLT (very large telescope). Dedicated software will be needed for
the co-addition of these tens of cubes (cumulative integration times that can reach
80 hours).
4.5.3 Mosaicking
In the mosaicking approach, a large number of (overlapping or not) sub-fields are obtained
and are then merged to create the final field of view. In the following, we will outline the
differences with the deep-field approach that has been described in detail in the previous
section and we will give examples of mosaics of 3D spectrographic fields.
! one or two observations are obtained per field of view. If two are obtained, a small
the deep-field approach, but for the offsets between the fields:
offset will be applied between them in order to facilitate the elimination of artifacts
! known offsets are applied between fields of view. Depending on the scientific goals
and to limit the introduction of systematic errors;
and on the observation constraints, these individual fields can be contiguous or not,
and the degree of overlap can vary.
The relatively small number of exposures per field of view compared to the case of
the deep field makes the problem of systematic errors much less critical. However, the
registration (see Section 4.5.2, Relative positioning) and normalization (see Section 4.5.2,
Relative normalization) steps are more troublesome in the case of a mosaic of fields.
60 60 60
3D spectrography
40 40 40
[NII]+Ha+[SII]
20 20 20 emission-lines
+
Arcsec
0 0 0
continuum
−20 −20 −20
Figure 4.16. Reconstructed images of NGC 2903 for a mosaic of 20 sub-fields obtained with
the OASIS/CFHT instrument.
! The number of spatial feature that could be used for registration purposes is very
! Any normalization would have to be propagated from one datacube to the next,
small and, quite often, they appear in a single sub-field (no overlap).
quite often using very small regions of overlap and yielding large errors.
A common solution is usually to make sure that a wide-field image in the same wavelength
range is available. It is then possible to systematically compare reconstructed images to
this wide-field image and derive the necessary registration and normalization information.
This allows for a single common reference for all the sub-fields but does not solve the
problems of the lack of reference sources for the registration in some sub-fields and the
problems of changes of observing conditions from one sub-field to another (heterogeneous
dataset).
Figure 4.17. Example of a mosaic of fields obtained using the OASIS/CFHT 3D spectrograph
on the active galaxy NGC 1068.
‘no-fill’ option of the XOasis ‘merge datacubes’ tool has been disabled, leading to the
creation of non-physical interpolated spaxels in the final datacube (and they are not that
easy to spot due to the quality of the interpolation).
4.6.2 Binning
Binning has the same objectives as smoothing and consists of putting together different
pixels/spaxels in order to increase the signal to noise in the data. The main difference
with smoothing is that it actually reduces the number of pixels/spaxels (there is no data
duplication). In the following, we will concentrate on spatial binning.
148 Pierre Ferruit
4
4050
2
2
V (km/s)
Arcsec
Arcsec
4000
0
0
−2
−2
3850
−4
−4
−4 −2 0 2 4 −4 −2 0 2 4 −4 −2 0 2 4
Arcsec Arcsec R (Arcsec)
3920 km/s 4000 km/s 4070 km/s 180 km/s 200 km/s 260 km/s
Figure 4.18. Example of kinematical maps reconstructed from a datacube binned using the
technique developed by Cappellari and Copin (2003) (courtesy of G. Dumas).
The main difficulty is the fact that the signal to noise ratio within a datacube is not
uniform. As a consequence, if binning is useful in some regions of a field of view (e.g. the
external regions of a galaxy) it may not be needed in others (e.g. the core of the galaxy
that would be present in the same observation). And, as spatial binning will decrease the
spatial resolution, binning may not only be useless in some regions but also unwelcome.
This has led to the emergence of the so-called adaptive binning techniques like the
one based on Voronoi tessellation developed by Cappellari and Copin (2003). These
techniques are called adaptive since the size of the bins is tuned in order to ensure the
signal to noise ratio in the binned datacube is as uniform as possible. This means that
spaxels with a high enough signal to noise ratio will remain untouched while the binning
will only be applied to areas where it is needed. The main constraints on the bins used
! The bin shall not overlap (this constraint is not satisfied by the smoothing techniques)
by Cappellari and Copin (2003) are the following:
! In order to try to avoid exotic bin shapes, the bins shall be as ‘compact’ (round) as
and no holes shall be introduced.
possible.
In their paper, these two authors present a method satisfying these constraints and its
practical implementation. It has been widely used in the SAURON project, and Fig-
ure 4.18 shows an example of maps reconstructed from an adaptively binned datacube.
! by the X-ray community for the automated detection of X-ray sources in surveys
have already been explored in various astronomical communities:
! by the radio community for the detection of sources in the HI surveys as an example.
(with a 3D touch as they can have energy-sensitive detectors).
surveys; see, for example, the SExtractor software from Bertin and Arnouts (1996).
Analysis of 3D data 149
Figure 4.19. Example of application of the crowded field spectro-photometry with sources (the
four quasar images) located on top of a non-uniform background (the lensing galaxy) (courtesy
of M. Roth).
In the optical and infrared 3D community, these techniques are only just starting to
emerge, and in the last few years work has been conducted by some members of the
Euro3D network (e.g. S. Foucaud and G. Covone). Two prototype tools have been devel-
oped: one based on the systematic use of SExtractor on monochromatic slices of a cube,
and the other based on the detection of correlations between spectra.
! How to efficiently and cleanly get rid of the backgrounds, which are usually not
imaging, the major issues are:
! How to efficiently collect the flux of an object that is spread over several spaxels and
uniform spatially?
! Using a single stellar template spectrum, a first guess of the stellar kinematics of the
is iterative and as follows:
! Using this first guess as a constraint (i.e. the kinematics is kept fixed during the fit),
object is derived.
! The first step is then repeated, this time using the optimal template, yielding a new,
of the ‘optimal’ template).
Figure 4.20. SAURON spectrum showing the superposition of the stellar continuum and of
emission lines (top panel) and the corresponding optimal template spectrum (linear combination
of stellar templates from a library) as determined using the pixel-fitting software developed by
J. Falcón-Barroso (courtesy of J. Falcón-Barroso).
Figure 4.21. Typical workflow for the fitting of the stellar continuum spectrum and of the
emission lines when they cannot be fitted simultaneously.
152 Pierre Ferruit
Image: specBN Row: #1
20
15
Pixel value
10
0
6300 6400 6500 6600 6700 6800
Position
Figure 4.22. Typical spectrum of the central region of NGC 2110 that will be used to prepare
the fit of the complete datacube. It includes lines from [OI] (doublet), [NII] (doublet), Hα and
[SII] (doublet).
! the existence of fixed or bounded line ratios (e.g. between lines of a same doublet);
these constraints will take into account:
! the fact that the width (specified here as the full-width-at-half-maximum, FWHM)
of a line shall be larger than the instrumental one;
Analysis of 3D data 153
Figure 4.23. Configuration file describing the fit to be conducted and the constraints that will
be used. Lines in the system are assigned numbers 1 to 7 (as they appear in the initial description
of the system).
! the fact that the velocities shall be within a few hundreds of km s−1 around the
The implementation of these constraints in the software is done through an ASCII con-
figuration file called the ‘stack’ file, which is displayed in Figure 4.23. At this stage, the
fit is performed with a single Gaussian profile.
Initial conditions
We also need to prepare a set of initial conditions for the fit. They will be stored in a
second ASCII configuration file called the ‘start’ file that is a mirror of the ‘stack’ file and
they can be obtained graphically in ‘fit/spec’. Note that in complex objects, it is quite
often necessary to prepare different initial conditions depending on the location within
the object.
Figure 4.24. Configuration file for a fit with two Gaussian profiles.
bars on the results from the fit. One possible way is to use the error estimate inferred
from the minimization itself (from how ‘sharp’ the minimum toward which the program
has converged is as a function of each fit parameter).
This functionality is not present in the ‘fit/spec’ software and we have therefore used
those of the spectra in the object. Note that it may be necessary to prepare several of
these template spectra if the spectral characteristics of the object vary significantly
! Create a set of noisy spectra with given signal to noise ratios by adding noise to the
within the field of view.
template. For each signal to noise ratio value, a large number of spectra (usually
! Automated fits are performed for all prepared spectra (i.e. 500 per signal to noise
500) are created in order to be able to perform a statistical analysis of the results.
! A statistical analysis is performed on the results to derive the accuracy of the results
value).
0.0
6300 6400 6500 6600 6700 6800
Position ( )
S/N = 3
S/N = 50
S/N = 5
S/N = 10
2
2300
1.8
2250
1.6
V_syst1
tmp
2200
1.4
2150 1.2
2100 1
0 0.5 1 1.5 2
logsn (unitless)
0 100 200 300 400 500
Sequence
Figure 4.25. Top panel: sketch illustrating how a Monté-Carlo type of approach is used to
determine the error bars. Bottom panels: fit of the results of Monté-Carlo approach to derive
an analytical relation giving the 1σ error bar on the centroid velocity of a line as a function of
the signal to noise in the input spectrum.
last contribution which is quite often not accounted for because it is extremely difficult
to estimate is the contribution of errors on the stellar continuum subtraction (so-called
template mismatch in particular).
Once all these steps have been performed the results are ready to be included in a
paper.
Table 4.1. Evolution of the total number of ‘pixels’ in a datacube for a set of
arbitrary selected 3D spectrographs.
Last, there are also discussions for some instruments to use an extremely conserva-
tive approach with respect to any resampling of the data and to keep working with
‘detector pixels’ until the very end of the processing/analysis process (the ‘ultimate’
solution). Although very interesting, this approach is extremely complex and requires
the development of specific analysis and visualization tools for every single step of the
data processing/analysis. This has not been implemented to date.
4.9 Conclusions
This course has presented a short but hopefully useful tour of various analysis methods
and strategies. The important message is to never forget to keep track of the noise and
systematic errors throughout the complete data analysis process.
If we make a critical analysis of the various methods that have been presented, one
relatively negative conclusion can be drawn: there are not enough real 3D tools. Most
of the tools presented correspond to generalizations of classical, already-existing tools
to a collection of spectra. They do not make use of the additional spatial information
usually present in the datacube (e.g. the fact that most of the time neighbouring spaxels
are correlated by the spatial PSF) and treat each spectrum independently of the others.
This clearly outlines that a significant effort has still to be made to take full advantage
of the 3D information contained in the datacubes.
Last, looking at how things will probably evolve within the next 5–10 years, there is a
huge challenge in front of us. This is illustrated by the numbers in Table 4.1 that show the
increase in size of the datasets over the years. Many of the analysis steps presented in this
document still rely heavily on a ‘manual’ handling of the data. This will not be possible
with data from, for example, the MUSE instrument. A huge effort will therefore have to
be made to prepare the automated tools that will be needed for the next generation of
3D spectrographs.
REFERENCES
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B., del Burgo, C., Fuensalida, J.J.
(1999), A&AS, 136, 189
Bender, R. (1990), A&A, 229, 441
Bertin, E., Arnouts, S. (1996), A&AS, 117, 393
Cappellari, M., Copin, Y. (2003), MNRAS, 342, 345
Falcón-Barroso, J., Sarzi, M., Bacon, R., Bureau, M., Cappellari, M., Davies,
R.L., Emsellem, E., Fathi, K., Krajnović, D., Kuntschner, H., McDermid, R.M.,
Peletier, R.F., de Zeeuw, T. (2006), New Astronomy Review, 49, 515
Ferruit, P., Mundell, C.G., Nagar, N.M., Emsellem, E., Pécontal, E., Wilson, A.S.,
Schinnerer, E. (2004), MNRAS, 352, 1180
Analysis of 3D data 157
Gibson, S.J., Taylor, A.R., Higgs, L.A., Brunt, C.M., Dewdney, P.E. (2005), ApJ, 626,
214
Hook, R., Fruchter, S. (2000), Astronomical Data Analysis Software and Systems IX,
ASPCS, Vol. 216, edited by Nadine Manset, Christian Veillet, and Dennis Crabtree
Kissler-Patig, M., Copin, Y., Ferruit, P., Pécontal-Rousset, A., Roth, M.M. (2003),
‘Euro3D data format, format definition’, http://www.aip.de/Euro3D, Euro3D data format:
Format definition, issue 1.2
Rousset (1992), ‘Contribution des méthodes numériques au dépouillement des données du spec-
trographe intégral de champ TIGER’, PhD thesis, Univ. J. Monnet de Saint-Etienne
Sánchez, S.F. (2004), AN, 325, 167
Wilman, R.J., Gerssen, J., Bower, R.G., Morris, S.L., Bacon, R., de Zeeuw, P.T.,
Davies, R.L. (2005), Natur, 436, 227
5. Science motivation for integral field
spectroscopy and Galactic studies
FRANK EISENHAUER
This lecture introduces the scientific motivation for integral field spectroscopy (IFS)
and describes the results from this novel technique in Galactic studies as of 2005. The
following chapter by Luis Colina then picks up on extragalactic studies and rounds out
the picture giving an outlook to future integral field spectroscopic studies. Following the
! science motivation;
five one-hour lectures, this chapter is broken down into sections on:
158
Science motivation for IFS and Galactic studies 159
kinematics, and line-strength distributions of nearby early-type galaxies’ (Bacon et al.,
2001a).
The development of a third generation of integral field spectrographs has recently
started. These instruments are built by large international consortia and are usually jus-
tified by a wide variety and mainstream science cases. Examples of this third generation
of instruments are MUSE (approximately 8.4 million euro, 172 staff years, 7 institutes)
and KMOS (approximately 6.5 million euro, 165 staff years, 6 institutes).
Presently, several integral field spectrographs are also being built or developed for
upcoming missions, for example PACS on the Herschel satellite and Near-Infrared Spec-
trograph (NIRSpec) on the James Webb Space Telescope (JWST).
! cover a very large field of view (larger than several square arcminutes);
objects that:
! are sparse and point-like (only few objects per several square arcminutes);
! require observations of a large spectral range at very high spectral resolution (larger
than several 10,000, covering a full atmospheric band).
! (all-sky) surveys, for example the Wisconsin Hα Mapper Northern Sky Survey is
Typical examples for unsuited science cases are:
done with fibre-fed (e.g. 2dF at the Anglo-Australian Telescope, AAT) or slit-mask
(e.g. Focal Reducer and Low Dispersion Spectrograph, FORS at the Very Large
! high spectral resolution (echelle) spectroscopy for planet search from Doppler velocity
Telescope, VLT) multi-object spectrographs;
measurement (e.g. High Accuracy Radial Velocity Planet Searcher, HARPS at the
European Southern Observatory, ESO, 3.6 m telescope) or the Lyman-alpha forest
observed in quasars (e.g. High Resolution Echelle Spectrometer, HIRES at the Keck
Telescope).
and substructure of ideal objects scale with distance. Therefore, the following list is
ordered by the distance, climbing the distance ladder by typically a factor 10 with each
rung.
The typical distance scale for the Solar System is one astronomical unit. An angular
separation of 1′′ corresponds to approximately 700 km, the ideal (1–10′′ ) field of view
(FoV) of an integral field spectrograph corresponds to 700–7000 km, e.g. the size of moons,
continents of planets and asteroids. The angular resolution (0.02–0.2′′ ) corresponds to
10–100 km, e.g. big volcanoes. Although many Solar System objects have already been
visited by spacecraft, there is still need for ground-based observations, for example to
obtain high-resolution spectra, and for monitoring long-term trends and seasons. Exam-
ples of Solar System studies taking advantage of integral field spectroscopy are ‘Martian
surface mineralogy from 0.8 to 1.05 µm TIGER spectra–imagery measurements in Terra
Sirenum and Tharsis Montes formation’ (Martin et al., 1996), ‘The three-dimensional dis-
tribution of Titan haze from near-infrared (NIR) integral field spectroscopy’ (Ádámkovics
et al., 2006) and ‘Near-Infrared Compositional Mapping of Vesta’ (Dumas et al., in
preparation).
The next step on the distance ladder leads to the most nearby stars at about 1 pc
(1′′ = 1 AU), for example Proxima Centauri. At that distance the ‘typical’ FoV covers
1–10 AU, therefore matching the size of planetary systems. The resolution is about 0.02–
0.2 AU, comparable to the separation of planets. However, present-day adaptive optics
and instruments do not provide enough contrast and no publications have emerged from
such observations yet.
The most nearby star-forming regions – for example the Taurus molecular cloud –
are located at a distance of about 100 pc (1′′ = 100 AU). The FoV of 100–1000 AU is
comparable to the size of circumstellar discs, and the angular resolution is 2–20 AU,
comparable to the orbital radius of Jupiter. So far, publications have concentrated on
prototypical stars, jets and their interaction with the ambient interstellar medium: ‘A
near-infrared spectral imaging study of T Tau’ (Herbst et al., 1996; Kasper et al., 2002),
‘Sub-arcsecond morphology and kinematics of the DG Tauri jet in the [OI]λ6300 line’
(Lavalley et al., 1997) and ‘The three-dimensional structure of HH 32 from GMOS IFU
spectroscopy’ (Beck et al., 2004).
The closest high-mass star formation regions, for example the Orion Nebula, are at a
distance of about 1 kpc (1′′ = 1000 AU). The FoV of 1000–10,000 AU and the resolution
of 20–200 AU is still appropriate to cover and resolve circumstellar discs and jets. The
publications address the processes involved in high-mass star formation, for example
‘Collimated molecular jets from high-mass young stars: IRAS 18151-1208’ (Davis et al.,
2004), and the impact of winds and radiation from the high-mass stars on circumstellar
discs of nearby stars, for example in ‘GEMINI multi-object spectrograph integral field
unit spectroscopy of the 167-317 (LV2) proplyd in Orion’ (Vasconcelos et al., 2005).
The Galactic Centre is located at a distance of about 10 kpc (10′′ ≈ 0.5 pc). At that
distance the FoV of 0.05–0.5 pc corresponds to the core size of a typical star cluster
and the resolution of 200–2000 AU is close to a light-day. Being a unique laboratory for
studying supermassive black holes and their environment, the Galactic Centre has been
the subject of a larger number of publications from integral field spectroscopy, for example
on ‘The nuclear cluster of the Milky Way: star formation and velocity dispersion in the
central 0.5 parsec’ (Krabbe et al., 1995) and ‘SINFONI in the Galactic Centre: young
stars and infrared flares in the central light month’ (Eisenhauer et al., 2005). A distance
of 10 kpc is also the typical scale for our Galaxy. Consequently, integral field observations
cover a wide variety of objects at this distance. Examples are ‘Integral field spectroscopy
162 Frank Eisenhauer
of faint halos of planetary nebulae’ (Monreal-Ibero et al., 2005) and ‘The X-ray binary
X2127+119 in M15: evidence for a very low-mass, stripped giant companion’ (Van Zyl
et al., 2004).
At a distance of 1 Mpc (1′′ ≈ 5 pc) we get to the closest disc galaxies, for example the
Andromeda galaxy M31. The typical FoV covers 5–50 pc, e.g. nuclear starburst regions,
and the spatial resolution is about 0.1–1 pc, the typical separation of stars in spiral arms.
Publications have so far concentrated on the nuclear regions, for example in ‘The M 31
double nucleus probed with OASIS. A natural vec m = 1 mode?’ (Bacon et al., 2001b).
The closest clusters of Galaxies – for example the Virgo cluster – are located at a
distance of about 10 Mpc (1′′ ≈ 50 pc). The FoV corresponds to about 50–500 pc, and
the resolution of 1–10 pc is still good enough to resolve individual starburst regions.
Again, publications focus mainly on the dynamics and properties of nuclear regions,
for example ‘The SAURON project: III. Integral-field absorption-line kinematics of 48
elliptical and lenticular galaxies’ (Emsellem et al., 2004), and ‘SINFONI adaptive optics
integral field spectroscopy of the Circinus galaxy’ (Müller-Sánchez et al., 2006).
Yet another factor 10 more distant (about 100 Mpc, 1′′ ≈ 0.5 kpc) are massive galaxy
clusters – for example the Coma Cluster – which host ‘extreme’ galaxies like ultra-
luminous infrared galaxies (ULIRGs). The field of view at that distance is approximately
0.5–5 kpc, about the size of galactic bulges. The resolution is about 10–100 pc. This is
sufficient to spatially resolve the velocity field of merging galaxies, a published example
is ‘Stellar dynamics and the implications on the merger evolution in NGC 6240’ (Tecza
et al., 2000).
The most distant objects observed so far with integral field spectrographs are at red-
shifts of z ≈ few (1′′ ≈ 5 kpc at z = 1). This early epoch – several Gyr after the Big
Bang – is particularly interesting because of its peak in cosmic star formation. The FoV
of about 5–50 kpc (at z = 1) covers one or more galaxies. The resolution of 0.1–1 kpc
allows us to resolve regions with sizes comparable to galactic bulges. The publications
focus on characterizing the physical properties of various galaxy samples, for example
‘Optical and near-infrared integral field spectroscopy of the SCUBA galaxy N2 850.4’
(Swinbank et al., 2005) and ‘SINFONI integral field spectroscopy of z ≈ 2 UV-selected
galaxies: rotation curves and metallicity gradients’ (Förster Schreiber et al., 2006), but
also cover various other topics like the feedback mechanism to the intergalactic medium,
for example in ‘Deep SAURON spectral imaging of the diffuse Lyman α halo LAB1 in
SSA 22’ (Bower et al., 2004).
The above examples illustrate the start of the race, and many more exciting results
are expected in the coming years. In parallel, several groups have started with the next-
generation integral field spectrographs and telescopes. The outlook for this future science
is given in the lecture by Luis Colina.
! ‘Imaging the universe in 3D’, Walnut Creek 1999 (Van Breugel and Bland-Hawthorn,
1994 (Comte and Marcelin, 1995): 40% Galactic versus 60% extragalactic;
! On the Galactic Centre: Eisenhauer et al. (2003a), ≈ 71; Eisenhauer et al. (2005),
a few citations. For the papers discussed in this lecture we find:
≈ 16; Genzel et al. (1996), ≈ 130; Genzel et al. (2000), ≈ 120; Genzel et al. (2003a),
! Related to star formation: Beck et al. (2004), ≈ 3; Davis et al. (2004), ≈ 7; Dougados
≈ 89; Krabbe et al. (1995), ≈ 178;
et al. (2003), ≈ 1; Herbst et al. (1996), ≈ 38; Kasper et al. (2002), ≈ 6; Lavalley et al.
However, it should be noted that this list is to a large extent limited to the ‘pioneering’
phase of IFS, and that the real impact is just about to start with the installation of
several facility instruments at large observatories.
Figure 5.1. Adaptive optics K-band (1.95–2.45 µm) image of the Galactic Centre region. The
field of view is approximately 20′′ × 20′′ . The arrows indicate the position of the black hole, which
is not visible in this image. Only during sporadic flares (every couple of hours) is the emission
strong enough to be detected. The angular resolution in the image is approximately 50 mas. For
comparison, the Schwarzschild radius of the Galactic Centre black hole is approximately 10 µas.
In other words, 3C273 emits the light of the whole Galaxy from a region comparable to
the size of the Solar System.
Subsequent observations of radio halos of quasars indicated that the energy source is
gravitation and not nuclear fusion: the energy contained in such radio halos is up to about
1054 J. If nuclear fusion were the energy source – for which the efficiency of converting
mass to radiation is of the order 1% – about 109 M⊙ would be required to power the radio
halos. However, the self-gravitational energy of this mass when squeezed into a couple of
light-days would exceed 1055 J, more than 10 times the fusion energy. Therefore, most of
the energy released in quasars must have a gravitational origin.
The evidence that the light source in quasars is as small as a black hole, however, was
weaker and more indirect: it was based on observations of radio jets. These jets reach
far out into intergalactic space and as such must ‘remember’ their ejection direction for
quite a long time. The orientation of the emitting region must thus be very stable, like a
‘gyroscope’. This indicates that the emitting object is probably spinning very fast. The
other key observation was that radio jets move close to the speed of light. This implies
indirectly that the speed of the spinning material is also close to that of light, i.e. it must
originate close to the Schwarzschild horizon of a black hole.
To summarize the epoch of 1970–90, most astronomers believed the supermassive black
hole picture. They found an enormous amount of indirect evidence, i.e. no other model
could reproduce the measurements equally well. But there has been no strict proof that
supermassive black holes exist, i.e. that the mass is enclosed within the Schwarzschild
radius, and that there is an event horizon. So the search for supermassive black holes
became a very hot topic, both in external galaxies (see lecture by Luis Colina) and in
our Milky Way (this lecture).
Excursion: line of argument for proving the Galactic Centre black hole
The line of arguments to prove the existence of the supermassive black hole in the Galactic
Centre is outlined in Figure 5.2. The first step is to measure its mass at ever-smaller radii.
Down to a few thousand Schwarzschild radii (RS ) this is accomplished by measuring
the orbits of stars with high angular resolution imaging (Figure 5.1) and integral field
spectroscopy (Figure 5.5). The combined astrometric and radial velocity observations also
provide the best measurement of the distance to the Galactic Centre, a prerequisite for
Science motivation for IFS and Galactic studies 165
Mass Radius
Velocity and
3 x 10 M Few thousand R
orbits of stars
Non-motion of Size of
radio source versus Few ten R radio
speed of stars source
Faintness, SED
Event horizon
Figure 5.2. Line of arguments to prove the existence of the supermassive black hole in the
Galactic Centre.
pinning down the absolute mass and radius. Already, these measurements exclude most
alternatives (for example a very dense cusp of stars or stellar remnants) to the black
hole hypothesis. Radio measurements of the size of Sgr A* (the name of the radio source
which is supposed to harbour the black hole) and the fact that the radio source hardly
moves when compared to the stars then make the case that the mass is indeed within
just a few tens of RS . The observed brightness variations and changes in the slope of the
spectra are interpreted to arise partly from orbital motions as close as a few RS . Finally,
it is the faintness of the Galactic Centre and the detailed shape of its spectral energy
distribution (Figure 5.6) that provide the theoretical argument that there is indeed a
black hole with a Schwarzschild horizon.
5.2.1 The nuclear cluster of the Milky Way: star formation and velocity dispersion
in the central 0.5 parsec
The first integral field spectroscopy observations of the Galactic Centre were published
by Krabbe et al. (1995). Their observations show that the radial velocity dispersion of 35
early- and late-type stars with distances of 1–12′′ from Sgr A* is 154 ± 19 km s−1 , which
strongly favours the presence of a central dark mass of 3 × 106 M⊙ within 0.14 pc of the
dynamic centre.
The observations were carried out in 1994 with the world’s first infrared integral field
spectrograph MPE 3D (Weitzel et al., 1996) visiting the ESO-MPI 2.2 m telescope in
La Silla. The spectra cover the wavelength range 1.9–2.4 µm at a spectral resolution
λ/δλ ≈ 1000. In total, 12 early-type emission line stars and 9 late-type stars with CO
band heads (see excursion) could be identified in the central 5′′ radius.
Figure 5.3. Typical spectra of early-type emission line stars (left) and late-type stars (right).
The figure shows three stars from the Galactic Centre region (courtesy of Genzel et al., 1996).
The emission lines from the early-type stars and the CO band-heads of late-type stars
allow the measurement of the radial velocity of the stars from the Doppler shift. The
measured radial velocities and projected distances from the Galactic Centre then give
the enclosed mass. From the virial theorem (see excursion) the estimate for the enclosed
mass within 0.14 pc is 4.0 × 106 M⊙ . The Bahcall and Tremaine (1981) estimator (see
excursion) for the enclosed mass within 0.19 pc is 6.3 × 106 M⊙ . When adding the veloci-
ties from the more distant stars, and also considering the gas dynamics at larger radii, the
gravitational potential in the Galactic Centre can be well fitted with the combination of
a 3 × 106 M⊙ central (black hole) mass and an isothermal stellar cluster with 3 × 105 M⊙
within a core radius of 0.5 pc (Figure 5.4). However, the data are still consistent with
a central mass dominated by a very compact cluster of stellar remnants (e.g. about 105
stellar black holes with a typical mass of 10 M⊙ ).
GM r
r̈ = − , (5.1)
r3
Science motivation for IFS and Galactic studies 167
Figure 5.4. Mass of the Galactic Centre black hole: the plots show the various estimates for
the encircled mass profile from the first integral field spectroscopic observations in 1995 to the
state of the art in 2005 (courtesy of Genzel et al., 1996; 2000; Krabbe et al., 1995).
where r is the vector from the point mass to the star, and G the gravitational constant.
If we then write
) *
d2 1
r · r = r · r̈ + ṙ · ṙ (5.2)
dt2 2
and use r · r = r2 , ṙ · ṙ = v 2 , and average <> over all stars, we get
1 d2 1
< r2 > = −GM < > + < v 2 > . (5.3)
2 dt2 r
Assuming that the ensemble of stars is in a statistical steady state, < r2 > is independent
d2 2
of time and dt 2 < r > = 0. If we further assume that the system is spherically symmetric,
then < v 2 > = 3 < vz2 > and < 1r > = π2 < R1 >, where vz is the radial velocity and R
the projected distance to the centre. From that, we get the virial estimator for the central
point mass:
3π < vz2 >
M= . (5.4)
2G < R1 >
5.2.2 The dark mass concentration in the central parsec of the Milky Way
In this second paper, Genzel et al. (1996) present an extended dataset, and improve their
mass estimate by considering a rotation component in the measured velocity dispersion
and by simultaneously fitting the observed surface density of stars.
The field of view covered in the observations is approximately 20′′ × 20′′ . Taking advan-
tage of additional high angular resolution speckle imaging data, the authors could extract
the spectra of 198 late-type stars, and 25 early-type stars. The projected stellar velocity
dispersion increases statistically very significantly from about 55 km s−1 at a projected
distance of 5 pc, to 180 km s−1 at 0.1 pc. When plotting the measured radial velocities
against the declination, a significant rotation of the stellar population is obvious. The
late-type stars follow general Galactic rotation, while the early-type stars show counter-
rotation. One therefore has to subtract (in quadrature) the rotation component
2vr2ot (R)
Mr ot = <R> (5.8)
πG
from the mass as derived from the virial theorem or the Bahcall–Tremaine estimator (see
previous section). vr ot (R) is the best-fit rotation curve.
The shortcoming of the virial theorem and Bahcall–Tremaine estimator is obvious: their
application assumes implicitly a homogeneous distribution of stars, but the measured
distribution is by no means homogeneous.
Therefore, Genzel et al. (1996) have also derived a less-biased enclosed mass esti-
mate from the first moment of the collisional Boltzmann equation, the ‘Jeans equation’
(for details see Binney and Tremaine, 1987). The problem is that the formula contains
the three-dimensional distance and the space density, and not the measured projected
distance and surface density. The transformation from volume to surface quantities is
mathematically described by Abel integrals. Since the deconvolution of the projected
quantities to the corresponding volume quantities is very sensitive to individual data
points and their errors, the authors parameterize the velocity dispersion and density dis-
tribution and find the best-fitting model to the data. The analysis gives a central dark
Science motivation for IFS and Galactic studies 169
6
mass of 2.5–3.2 × 10 M⊙ at 6–8 σ significance (Figure 5.4). The data also show that
the mass-to-light ratio significantly increases towards the Galactic Centre up to at least
100 L ⊙,M2 ⊙µm , indicating the ‘dark’ nature of the central mass. It is therefore either a com-
pact cluster of 105 stellar black holes with a typical mass of 10–20 M⊙ , or a single massive
black hole.
5.2.3 Stellar dynamics in the Galactic Centre: proper motions and anisotropy
This paper by Genzel et al. (2000) combines proper motion data from speckle interfer-
ometric imaging with radial velocity measurements from integral field spectroscopy in
a new analysis of the stellar dynamics in the Galactic Centre. They find that the sky-
projected velocity components of the young, early-type stars indicate significant devia-
tions from isotropy, with a strong radial dependence. Most of the bright emission line
stars at separations from 1′′ to 10′′ from Sgr A* are on tangential orbits. This tangential
anisotropy of the He I stars and most of the brighter members of the IRS16 complex
is largely caused by a clockwise (on the sky) and counter-rotating (line of sight, com-
pared to the Galaxy), coherent rotation pattern. The assumption of velocity isotropy
in the above mass estimates is clearly violated. The authors thus explicitly include the
velocity anisotropy in estimating the central mass distribution. This Leonard and Mer-
ritt (1989) mass estimator (see excursion) confirms previous conclusions that a compact
central mass concentration (central density ≥ 1012.6 M⊙ / pc3 ) is present and dominates
the potential at 0.01–1 pc. Depending on the modelling method used, the derived central
mass is 2.6–3.3 × 106 M⊙ .
Figure 5.5. Three-dimensional orbits and distance determination for the Galactic Centre: fit-
ting simultaneously the measured radial velocities and motion on the sky gives directly the
distance to the Galactic Centre. The radial velocity also removes the ambiguity in the inclina-
tion when fitting the stellar orbits solely from their motion on the sky (courtesy of Eisenhauer
et al., 2005).
of the central point mass requires the distance to the Galactic Centre. This distance
is known from various other measurements to be 8.0 kpc, but only to an accuracy of
±0.5 kpc. Since the mass estimate scales with the third power of distance, the distance
is also the major uncertainty (about 20%) in the mass determination.
5.2.5 SINFONI in the Galactic Centre: young stars and infrared flares
in the central light month
This paper by Eisenhauer et al. (2005) focuses on the stellar population within the central
light-month and flares from the black hole (see below), but also gives an update on the
mass and distance to the Galactic Centre. The paper is based on the first adaptive optics-
assisted integral field spectroscopy observations of the Galactic Centre with SINFONI
(Eisenhauer et al., 2003b; Bonnet et al., 2004), which allowed measurement of the radial
velocities of about a dozen stars in the central arcsecond, and to add another epoch to
the S2 orbit (Figure 5.5). The updated estimate of the distance to the Galactic Centre
from the S2 orbit fit is 7.62 ± 0.32 kpc, resulting in a central mass value of 3.61 ± 0.32 ×
106 M⊙ (Figure 5.4).
Science motivation for IFS and Galactic studies 171
Figure 5.6. Spectral energy distribution (SED) of the Galactic Centre. The measurements of
the slope at near-infrared frequencies strengthen the radiative accretion models and as such
provide further evidence for an event horizon (courtesy of Eisenhauer et al., 2005).
5.2.6 Again: SINFONI in the Galactic Centre: young stars and infrared flares
in the central light month
This paper by Eisenhauer et al. (2005) was already mentioned for the mass and distance
determination. The authors also report the first measurements of the slope of the spec-
tral energy distribution (SED) in the infrared during a flare. They find that the SED at
′
1.7–2.45 µm is fitted by a featureless ‘red’ power law Sv ∝ v −α of spectral index −4 ± 1
(Figure 5.6). The observed spectral slope strengthens synchrotron models in which the
172 Frank Eisenhauer
infrared emission comes from accelerated, non-thermal, high-energy electrons in a radia-
tively inefficient accretion flow.
This support for the radiatively inefficient accretion is a step towards ‘proving’ the
existence of the event horizon of the Galactic Centre black hole. The main evidence for the
presence of such an event horizon – i.e. that matter, energy and radiation cannot escape
from inside this region – is the faintness of the Galactic Centre supermassive object.
It is several thousand times fainter than what is expected from the mass–luminosity
relation for the Centre of other galaxies (Nagar et al., 2005) and about a million times
smaller than the Eddington luminosity (see excursion). This very low luminosity is best
explained by radiatively inefficient accretion models (see excursion), which inherently
rely on the presence of an event horizon. To ‘prove’ the existence of the event horizon
the observations must constrain the parameters of the accretion models. The strongest
observational constraint for the accretion models – and thus for the presence of an event
horizon – is the SED of the Galactic Centre. In the energy range from radio to X-
rays, the main features of the SED are the synchrotron radiation at radio wavelengths
rising with a power law up to about 1012 Hz, a maximum at sub-millimetre wavelengths,
a strong decline in the infrared, and strong emission at X-ray energies during flares.
The connection between the infrared and X-ray emission is particularly important in
constraining the accretion models – and thus the event horizon hypothesis – because
it is very sensitive to the physical parameters of the emitting region, e.g. size, density,
temperature and magnetic field strength. Coordinated multi-wavelength observations and
infrared integral field spectroscopic observations are targeting this topic.
σT F
FR = (5.11)
c
) 2 *2
8π e
σT = (5.12)
3 me c2
L
F = (5.13)
4πr2
where σT is the Thompson cross-section for electrons (mass me , charge e, speed of light
c) and F is the radiation flux at the distance r for given luminosity L.
The gravitational force on the plasma from the central black hole is dominated by the
much more massive protons:
GM mP
FG = , (5.14)
r2
where G is the gravitational constant, mP the mass of a proton and M the mass of
the black hole. To maintain a stationary accretion flow in the plasma, gravitation must
Science motivation for IFS and Galactic studies 173
dominate over the radiation pressure (FG > FR ). This leads to the Eddington limit:
4πGmP c M
L < LEddington = = 1.3 × 1038 erg s−1 . (5.15)
σT M⊙
5.3.1 The nuclear cluster of the Milky Way: star formation and velocity dispersion
in the central 0.5 parsec
This paper, by Krabbe et al. (1995), showed that the central parsec of the Galaxy is
powered by a cluster of about two dozen luminous hot stars.
The authors could identify two main groups of emission line stars and model their
spectra with non-LTE (local thermal equilibrium) stellar atmospheres with winds: the
first group are the so-called ‘He-stars’, which show strong emission in He I and Brγ (see
also Figure 5.3). These stars are classified as WN9 (late nitrogen-rich Wolf–Rayet stars)/
Ofpe (O stars with peculiar emission lines) stars with temperatures of 20,000–30,000 K.
The emission lines show P Cygni profiles, indicating winds with a mass loss rate of
> 10−5 M⊙ yr−1 . Several stars are found to have a very large He I/Brγ ratio and seem
to be pure He-stars. The second group of emission lines stars are of type WC (carbon
>
rich Wolf–Rayet stars) with temperatures ∼ 30,000 K. These stars show C III and N III
emission lines with a width of several 1000 km s−1 , indicating very strong winds. The
174 Frank Eisenhauer
progenitor stars must have had zero age main sequence (ZAMS) masses up to about
100 M⊙ . The hot star cluster can fully account for bolometric and ionizing luminosity of
central parsec of the Galaxy.
The authors could also identify more than a dozen late-type stars by the strong CO
band-head absorption in their spectra. They found four supergiants with spectral type
K/M. These stars have K-band luminosities 10,000 times larger than the Sun. There
are also more than 10 late-type giants within the central 8′′ , maybe so-called asymptotic
giant branch (AGB) stars.
Putting together the findings from their integral field spectroscopy, the authors can
model the stellar population and star-forming history (see excursion on stellar popula-
tion modelling). The following observations have to be reproduced by the model: Lb ol ≈
N (KM I)
107 L⊙ , LLLb yo cl ≈ 10, LLLKy c < 0.1, N (OB + WR/HeI with Lb ol > 5.5) ≈ 3, N (OB+W R/HeI) ≈
N (OB)
0.1 and N (W R/HeI) < 0.1.
Even without detailed modelling, some scenarios are obviously excluded; for example,
continuous star formation – i.e. constant star formation over more than Gyrs – cannot
account for the high number of Wolf–Rayet stars. If one uses our Galaxy as a template
for continuous star formation – where we observe only of the order 100 Wolf–Rayet stars
for a total stellar mass of 2 × 1011 M⊙ – one would expect no (< 0.001) WR star in
the central 10′′ with its < 107 M⊙ , but several such stars are observed. Also, constant
star formation – i.e. star formation that started some time ago and has continued at a
constant rate since then – can be excluded: the large LLLb yo cl > 10 would require that the
star formation has started more than 107 years ago. This would imply that the first late-
N (KM I)
type supergiants had formed, so that N (OB+W R/HeI) > 1. However, the observed ratio is
less than 0.1.
The best-fitting model is a small starburst approximately 7 ± 1 Myr ago, with a decay
time of 3–4 Myr. About 103.5 stars have formed in this small starburst. In this scenario
the Galactic Centre stars are presently in a short-lived, post main sequence wind phase.
The presence of the intermediate mass late-type stars cannot be explained by that
scenario and indicates another star formation event about 108 years ago. Another problem
with the small starburst model is that it predicts about 15 early-type (OB) main sequence
stars, which have not been found in the data of Krabbe et al. (1995). However, this non-
detection is solely based on the sensitivity limits at that time, and these stars have been
found with the next generation integral field spectrograph SINFONI at the VLT.
5.3.2 The dark mass concentration in the central parsec of the Milky Way
In this second paper, Genzel et al. (1996) also analyse in detail the spatial distribution
and dynamics of the central star cluster. In particular, they find that the early-type stars
(He stars, WR/Of stars) are strongly concentrated within a radius of 12′′ , and that these
stars show a large and coherent counter-rotation. This strongly supports their origin in a
well-defined epoch of star formation (see above), presumably when a dense cloud fell into
Science motivation for IFS and Galactic studies 175
the central parsec less than 10 Myr ago. Apart from the early-type stars, the late-type
stars (supergiants and bright AGB stars) form a ring at 4–10′′ . The authors find that
the most likely explanation for the absence of the largest red stars (with a stellar radius
of ≈ 50–900 R⊙ ) in the ≈ 5′′ central ‘hole’ is their destruction by collisions with main
sequence, solar-type stars.
5.3.3 Stellar dynamics in the Galactic Centre: proper motions and anisotropy
As described in the discussion on the black hole mass, Genzel et al. (2000) find that many
of the emission line stars at 1–10′′ radius not only show rotation in the radial velocity but
also rotate clockwise on the sky. This supports even more strongly the notion that the
rotation of the young star cluster may be a remnant of the original angular momentum
176 Frank Eisenhauer
Figure 5.7. Two stellar discs in the Galactic Centre. Left: measured velocities in the clockwise
disc and perpendicular to it (courtesy of Levin and Beloborodov, 2003). Right: the same for the
counter-rotating disc (courtesy of Genzel et al., 2003a).
pattern in the interstellar cloud from which these stars were formed. The authors also
report the first proper motions of the ‘S-stars’ within the central 1′′ . The absence of CO
emission in the integrated spectrum of that region indicates that these S-stars are also
early-type stars. If they are on the main sequence, they would be of type B0 to B2. As
most of the S-star cluster members are also on clockwise orbits, the authors propose that
the S-stars are those members of the early-type cluster that happen to have low angular
momentum and can thus plunge into the immediate vicinity of Sgr A*.
5.3.4 Stellar disk in the Galactic Centre: a remnant of a dense accretion disk? and
the stellar cusp around the supermassive black hole in the Galactic Centre
Levin and Beloborodov (2003) and Genzel et al. (2003a) were the first to interpret the
observed rotation of the early-type stars in the central 10′′ as a thin stellar disc.
Levin and Beloborodov (2003) find that out of 13 stars whose three-dimensional velo-
cities have been measured by Genzel et al. (2000), 10 lie in a thin disc. If the stars are
rotating in a common disc, the velocity vectors vi of all stars must be perpendicular to
the normal vector, n, of the disc. The authors thus search for the vector n that minimizes
(n · vi )2
N
1 $
χ2 = , (5.16)
N − 1 i=1 (nx σxi )2 + (ny σy i )2 + (nz σz i )2
where nx,y ,z are the components of the normal vector and σxi,y i,z i the errors in the
measured velocity components of the stars. N is the number of stars. Figure 5.7 shows the
velocities of the stars parallel and perpendicular to the disc for the best-fitting orientation.
Simulating an artificial sample of flared discs, the authors deduce that the half-opening
angle of the disc is consistent with zero within the measurement errors, and does not
exceed 10◦ .
Levin and Beloborodov (2003) propose that a recent burst of star formation has
occurred in a dense gaseous disc around Sgr A*. In such a self-gravitating disc the parti-
cle density is at least four orders of magnitude larger than that of the densest molecular
clouds in the Galaxy, enough to overcome the tidal forces to start fragmentation. The
initial characteristic mass (Jeans mass) is accordingly high, about 1.8 M⊙ . Therefore, star
formation in a disc could naturally explain the observed dominance of high-mass stars.
Taking advantage of the adaptive optics at the VLT, Genzel et al. (2003a) could signifi-
cantly extend the sample of stars with proper motions. Following Levin and Beloborodov
(2003), Genzel et al. (2003a) find a second disc of stars (Figure 5.7). The two discs are
Science motivation for IFS and Galactic studies 177
1.05
HeI 2.1127
293 (±35) km/s
1.3(±0.3) Å
1.00
Sν ν (Normalized)
HI Brγ 2.1661
464 (±20) km/s
0.95 5.8(±0.4) Å
-2
0.90
<1Å
0.85 FWHM=
1200(±80) km/s
EW=11.7(±2) Å
Wavelength (µm)
Figure 5.8. The ‘S-star’ cluster in the central arcsecond of the Galactic Centre. Left: integrated
SINFONI K-band (1.95–2.45 µm datacube with spectral classification of the S-stars: solid cir-
cles indicate early type stars, dashed circles indicate late-type stars). Right: average spectrum
of the bright (mK < 15, top) and faint (mK = 15 − 16, bottom) early-type stars (courtesy of
Eisenhauer et al., 2005).
inclined at large angles and counter-rotate with respect to each other. The authors also
find that the proper motions of the many more stars without radial velocities are not
randomly orientated but group at maximum absolute projected angular moment along
the line of sight:
Jz xvy − yvx
=. (5.17)
Jz,max (x2 + y 2 )(vx2 + vy2 )
where x, y and vx , vy are the sky coordinates and velocity components. Jz /Jz ,m ax =
+1, 0, −1 for clockwise tangential, radial, counter-clockwise tangential orbits. The group-
ing of the stars in absolute projected angular moment fits the two discs well.
Based on the spectra from integral field spectroscopy (see above), the stellar content
of the two discs is essentially the same, indicating that they formed at the same time.
Although other scenarios cannot yet be fully excluded, the ‘paradox of youth’ on a parsec
scale may be solved: the authors conclude that of the possible formation scenarios for
these massive stars the most probable one is that 5–8 Myr ago two clouds fell into the
centre, collided, were shock-compressed and then formed two rotating (accretion) discs,
which then fragmented and formed the stars.
5.3.5 SINFONI in the Galactic Centre: young stars and infrared flares
in the central light month
This paper by Eisenhauer et al. (2005) reports on 75 mas resolution, NIR imaging spec-
troscopy within the central arcsecond of the Galactic Centre. The authors find that to a
limiting magnitude of K ≈ 16, 9 out of 10 stars in the central 0.4′′ , and 13 out of 17 stars
out to 0.7′′ from the central black hole have spectral properties of B0–B9 main sequence
stars (Figure 5.8). The spectral classification is based on the detection of the Brγ and
He I lines. Different from the emission line stars in the stellar discs at 1–10′′ , the S-stars
show Brγ in absorption, which is characteristic of B-type stars. Another indication is the
lack of the typical CO band-head absorption of late-type stars (see Figure 5.3). Averag-
ing the spectra of the brighter and fainter stars, the signal-to-noise ratio is large enough
even to distinguish the spectral sub-type from the equivalent width (EW) of the He I
line. The bright stars (mK < 15) show strong (EW > 1 Å) absorption, typical for B0–B3
178 Frank Eisenhauer
stars, while the fainter (mK = 15–16) stars have very little H eI absorption (EW < 1 Å),
typical of B5–B9 stars. The low K-band luminosity and the detailed shape of the Brγ
line indicate that the S-stars are main sequence and not giants or supergiants.
The fact that nine out of 10 S-stars are early-type (= young) stars makes the ‘paradox
of youth’ even stronger. A possible explanation could be stellar mergers. Such mergers
would probably spin up the stars. However, based on the He I line width, all brighter
S-stars have normal rotation velocities of about 150 km s−1 , similar to solar neighbour-
hood stars. Therefore, the authors rate the stellar merger scenario as less likely.
As for the emission line stars at 1–10′′ , the explanation for the ‘paradox of youth’ could
come from the dynamical properties of these stars. Combining the radial velocities from
the integral field spectroscopy with imaging data, the authors could derive improved
three-dimensional stellar orbits for six of these S-stars. The orientations of the orbits
appear random and the orbital planes are not co-aligned with those of the two discs of
massive young stars 1–10′′ from Sgr A*. The authors thus exclude the hypothesis that
the S-stars as a group inhabit the inner regions of these discs.
But have the S-stars ever been in a disc? One mechanism to change the direction of the
angular momentum is Lens–Thirring precession from the rotating black hole. The rotating
space time breaks the spherical symmetry, which would normally preserve the angular
momentum. The period of this Lens–Thirring precession is (e.g. Levin and Beloborodov,
2003):
) *2 ) *3/2 ) *3
6 3 × 106 M⊙ 1 − e2 r0
PL T = 5.8 × 10 a −1
yr, (5.18)
M 1 − 0.872 4.6 × 10−3 pc
where a is the dimensionless spin parameter (a = 1 for maximum spin), M the mass of
the central black hole, and r0 and e the semi-major axis and eccentricity of the orbit.
For a spin parameter of a = 0.52 (Genzel et al., 2003b), however, the lifetime (10 Myr)
of two out of the six B-stars is too short for Lens–Thirring precession to account for the
change in angular momentum. Therefore, other mechanisms are necessary to explain the
S-stars (see also the excursion on the ‘paradox of youth’).
One potential explanation is a dissolving star cluster with an intermediate mass black
hole (IMBH), possibly formed by runaway merger in the cluster core. But it is not clear
if an IMBH can have such a tightly bound cluster, and the IMBH would have to be two
orders of magnitude more massive than can be explained by formation from a runaway
merger. Another explanation might be a massive binary exchange, where the S-stars
were originally in a massive (≈ 100 M⊙ ) binary system and switched partner in a close
interaction with the black hole. However, there is no evidence of such massive WR/O stars
on highly eccentric orbits, and more massive stars should also have been captured, which
is not observed. The last scenario discussed here is exchange capture with stellar black
holes, in which the B-stars play billiards with equal mass stellar black holes in a central
cusp. The advantage of this scenario is that it would favour the observed eccentric orbits
and would prefer B-stars because they best match the mass of the about 10 M⊙ stellar
black holes. This scenario is also supported by X-ray observations indicating the presence
of about 104 neutron stars and stellar black holes in the Galactic Centre region. However,
this scenario requires a large reservoir of B-stars at larger radii, which are not found.
In summary, the ‘paradox of youth’ for the S-stars has not yet been solved, although the
observations indicate that these stars are probably brought into the central light-month
by strong individual scattering events.
Science motivation for IFS and Galactic studies 179
a)
b)
d)
c)
Figure 5.9. The four stages of star formation: a) cores form within molecular clouds; b) pro-
tostar with surrounding nebular discs; c) stellar wind creating bipolar outflow; d) newly formed
star with circumstellar disc (courtesy of Shu et al., 1987).
1
http://www.mpia-hd.mpg.de/homes/dullemon/lectures/starplanet/
180 Frank Eisenhauer
Figure 5.10. The formation of discs and outflows: a,b) gravitational collapse of a magnetically
supported molecular cloud; c) formation of an accretion disc; d) magnetically threaded disc;
e) bipolar outflow; f) head of the jet (courtesy of Cornelis Dullemond).
a) 1.4
1.3
T Tau NW K
1.2
1.1
1.0
0.9
0.8
0.7
2 2.1 2.2 2.3 2.4
1
b) c) 6 v=0 v=1
0 TTau NW 4
1-0 S(0)
5
1-0 S(1)
3
1-0 S(2)
1-0 S(7)
2
1
2-1 S(3)
2-1 S(2)
4 0
7
3
T = 2500
6
H2
2-1 S(1)
T = 2000 2 5
T = 1500
1 4
3
2
0 1
0 2000 4000 6000 8000 10000 0
Figure 5.11. Emission from molecular hydrogen in T Tau: a) K-band spectrum of T Tau NW.
The flux ratio of H2 ν = 1 − 0 S(1) and H2 ν = 2 − 1 S(1) (highlighted) give a simple first estimate
for the excitation mechanism; b) excitation diagram (both figures courtesy of Herbst et al., 1996);
c) level diagram with H2 ν = 1 − 0 S(1) transition.
of T Tau NW and the west filament suggests that one star of the binary system produces
each outflow. Indirect evidence suggests that T Tau NW is associated with the primary
star, and the west filament with the infrared companion. The thermal, diffuse H2 emission
observed at the location of the stars could be caused either by collimation of the NW
and W outflows or by accretion shocks, which would also explain the large extinction
and infrared luminosity of the companion.
3.3
T Tau N 3.1
2.9
2.7
0.69'' 1.45
1.35
12CO 1.25
T Tau S
Br10
T Tau S 1.15
1.05
Mg Na Ca Mg
0.95
2.05 2.15 2.25 2.35
Wavelength [µm]
Figure 5.12. Integral field spectroscopy of T Tau. Left: Br10 (1.736 µm) image; right: K-band
spectrum of T Tau N (top) and T Tau S (bottom) (courtesy of Kasper et al., 2002).
respective emitting regions. All this indicates that T Tau S is not stellar light reddened by
strong absorption from cold dust, but that the emission arises from hot dust. Therefore
a protostar with an infalling envelope seems an unlikely explanation for T Tau S. On the
other hand, the results are consistent with a model that describes T Tau S as a pre-main
sequence star surrounded by a small edge-on disc. The disc is observed at an inclination
such that it obscures the star, at the same time leaving the line of sight towards the
Bracket emission from the polar regions relatively unobscured.
! What is driving the jets? There are a number of competing explanations (see excur-
There are two long-standing questions on microjets:
sion on star formation): disc winds, x-winds and stellar winds. Unfortunately, the
available magnetohydrodynamical models do not predict line ratios, and simple spec-
troscopy cannot answer the question. Instead, the models predict velocities, densities
and magnetic field strengths. We thus have to measure quantities like diameter, den-
sity, terminal velocity and rotation speed to identify the driving mechanism of the
! What is the heating mechanism? Given the small launching radius (0.1–10 AU) of the
jets.
jets, dramatic adiabatic cooling should occur in the expansion region, and extended
forbidden line emission should not be detectable. This is in disagreement with the
observations. Several heating processes are discussed for the microjets: shocks against
the environment or within the jet, turbulence in a viscous mixing layer, ambipolar
diffusion heating and compression by jet instabilities.
Several integral field spectroscopic studies have been aiming at resolving these questions.
Science motivation for IFS and Galactic studies 185
Disk Wind (Macc = 10−5 Msun yr−1) convolved with 14 AU beam
Vshift(km/s)
B −0
50 CW Tau −20
HH 30
HL Tau −40
RW Aur
A
Vshift(km/s)
0
−20
0
0 50 100 150 200
−40
Projected distance from source (AU)
80 60 40 20
Distance R across the jet (AU)
400 [O I] RW Aur
Projected Vrad (km/s)
200
0
−200
−400 Sol A α = 0.003
−200 −100 0 100 200
Projected distance along jet (AU)
Figure 5.13. Comparison of the RW Aur microjet with cold disc wind models. Top left: jet full-
width-at-half-maximum (FWHM) measured from the [SII]λ6731 line; right: transverse velocity
shifts between symmetric positions on either side of the jet axis (both figures courtesy of Douga-
dos et al., 2003); bottom: position–velocity diagram measuring the terminal velocity (courtesy
of Garcı́a, 2006).
Position ('')
2
6
0 2 4 0 500
R.A. offset (arcsec) Velocity (km/s)
[OI] [Ha] [N II] [SII]
2 HV
0
Offset (arcsec)
2 IV
2 LV
0 2 4 0 2 4 0 2 4 0 2 4
Offset (arcsec)
Figure 5.14. DG Tauri jet. Top left: deconvolved [O I]λ6300 map with an effective resolution
of 0.35′′ ; top right: position–velocity diagram along the jet axis, symbols indicating the centroid
velocities of the multi-component line fit (both figures courtesy of Lavalley et al., 1997); bottom:
channel maps for [O I]λ6300, Hα, [N II]λ6583 and [S II]λλ6716,6731 lines at [−400, −250] km s−1
(top), [−250, −100] km s−1 (middle), [−100, +10] km s−1 (bottom) (courtesy of Lavalley-Fouquet
et al., 2000).
The observations were obtained with the Tiger and OASIS instruments at the CFHT
in 1994 and 1998 and cover the [O I]λ6300, Hα, [N II]λ6583 and [S II]λλ6716,6731 lines
(the numbers following λ indicate the wavelength in Å). The morphological structure
of the mass outflow is revealed with unprecedented spatial resolution of approximately
0.35–0.5′′ . The deconvolved [O I] map (Figure 5.14 top left) shows an unresolved inner
peak containing two thirds of the total line flux, followed by a collimated jet-like body
extending out to ≈ 1.5′′ from the star, two resolved knots at distances of 2.7′′ and 4′′ ,
and a counter-jet. The unresolved inner peak is slightly displaced (≈ 20 AU) from the
star. It is apparently stationary and possibly linked to the initial jet collimation. The
size limit and absolute photometry for this peak allows Lavalley et al. (1997) to set
new constraints on the jet mass loss rate of 1.3 × 10−8 – 6.5 × 10−6 M⊙ yr−1 , lowering
previous estimates by a factor 5 to 25. From the [O I], Hα, [N II] and [S II] channel maps at
different velocities (Figure 5.14 bottom), the fast jet core appears surrounded by a slower-
moving flow or cavity. The two velocity components are analysed in more detail from the
position–velocity diagram (Figure 5.14 top right). A multiple component line fit for each
position clearly shows the two distinct velocity components at distances less than 1.5′′ :
the high-velocity component quickly accelerates up to 350 km s−1 ; the slow component
Science motivation for IFS and Galactic studies 187
follows the fast component at proportionally slower velocities up to about 100 km s−1 .
Accordingly, the slow component is interpreted as circumstellar material entrained in the
jet. Beyond 1.5′′ (200 AU) only one velocity component is detected, and the jet starts
wiggling around (Figure 5.14 top left and bottom). The knot at 2.6′′ shows a curved
morphology (Figure 5.14 top left). This and the transverse velocity gradient are strongly
suggestive of a resolved bow shock. Radial velocity jumps of ≈ 50 km s−1 are observed
between the bright knots at 1′′ and 3.6′′ . In order to determine the heating process of
the jet, Lavalley-Fouquet et al. (2000) also compare the measured [N II]λ6583/[O I]λ6300
and [S II]λ6716/[S II]λ6731 versus [S II]λ6731/[O I]λ6300 line ratios with available model
predictions for shocks, ambipolar diffusion and turbulent mixing: the measurements are
best reproduced by shocks with speeds of ≈ 50–100 km s−1 and pre-shock densities of
≈ 105 –103 cm−3 .
Combining the evidence – the bow shape structure of the knot, the velocity jumps at
1–4′′ , the wiggling beyond 1.5′′ , the line ratios and the variability in the ejection velocity
reported elsewhere – indicates that ‘shocks’ is the answer to the question of what is
heating the microjets.
1.0"
Figure 5.15. Herbig–Haro object HH 32 A. Left: observed Hα radial velocity channel maps;
right: channel maps computed from a 3/2-dimensional bow shock model (courtesy of Beck et al.,
2004).
r
z
zmax vflow
vf low
vshock
the model does not show the same complexity as the data, and in particular it fails to
reproduce the line ratios. The reason for the failure is that the parameterization is much
too simple for the spatially complex region. Another shortcoming of the 3/2-dimensional
model is that it treats the shock as locally planar, i.e. by assuming that post-bow shock
cooling distance is small compared to the bow shock. In HH 32 A, however, the cooling
distance is comparable to the size of knots.
In summary, the integral field spectroscopic observations of HH 32 A reveal two or
three bow shocks along the redshifted body of the HH 32 outflow.
Figure 5.16 illustrates the geometry of the 3/2-dimensional model. The result from such
a 3/2-dimensional model for the Herbig–Haro object HH 32 A is shown in Figure 5.15.
Science motivation for IFS and Galactic studies 189
Excursion: line ratio diagnostics in the optical
The optical wavelength range is rich in atomic lines caused by transitions between elec-
tronically excited states. The strength of an emission line is sensitive to the physical
conditions in the emitting region. Therefore line ratios provide a diagnostic tool for
deriving temperature, densities and other properties. A good introduction to this topic
can be found in Osterbrock (1989). Here are a few examples of diagnostic line ratios used
! [O I I I ]λλ(4959+ 5007) and [N I I ]λλ(6548+ 6583) : [O III] and [N II] are amongst the best
in the analysis of stellar jets and Herbig–Haro objects:
[O I I I ]λ4363 [N I I ]λ5754
examples of ions that have emission lines from two different upper levels with signif-
icant different excitation energies and for which both lines are easily observable in a
single spectrum. For low densities < 105 cm−3 , where collisional excitation is negligi-
ble, the level population follows the Boltzmann formula, and the line ratio traces the
excitation temperature (larger ratio indicating lower temperature). At high densities
> 105 cm−3 collisional de-excitation begins to play a role. Because the lower excita-
tion state has a considerably longer lifetime – a lower radiative transition probability
– it can be more quickly collisionally de-excited, thus weakening its line emission. In
! [O I I ]λ3729 and [S I I ]λ6716 : these line ratios are examples of density tracers. Because
this case the line ratio traces the density (higher ratio indicting lower density);
[O I I ]λ3726 [S I I ]λ6731
the two lines are emitted from levels with nearly the same excitation energy, the exci-
tation rate is almost independent of temperature. The relative strength of the opti-
cal transitions is set by the competing collisional de-excitation, which de-populates
the state more efficiently with the longer lifetime, i.e. a smaller radiative transition
probability (see above). Because this collisional de-excitation is strongly enhanced at
higher densities, the above line ratios are tracing directly the density of the plasma.
! [N I I ]λ6583 and [O I ]λ6300 : if the emission lines of the above indicators are too weak
A high line ratio indicates a low electron density;
Hα Hα
to be observed, line ratios from different species and ionization states can be used
to determine the physical parameters of the plasma; for example, the ratio of lines
with high excitation energies (e.g. [N II]λ6583 and [O I]λ6300) and the low excitation
Hα line trace the excitation temperature of the plasma. A high line ratio indicates
a high excitation temperature.
5.4.6 Collimated molecular jets from high-mass young stars: IRAS 18151−1208
This paper by Davis et al. (2004) presents integral field spectroscopy of the high-mass
protostar IRAS 18151−1208, its jet and the associated shocks. IRAS 18151−1208 is a
class I young stellar object with a double-peaked spectral energy distribution dominated
190 Frank Eisenhauer
0.3
0.2
Flux (Jy)
0.1
0
0.008 Brγ co 3–1
2–0 4–2
5–3
6–4
7–5
0.004
H2 0
1.6 1.8 2 2.2 2.4
Wavelength (µm)
0.008
1–0S(1)
Q–branch
0.007 Knot/Region B3
0.006
0.005
Flux (Jy)
0.004
1–0S(0)
1–0S(2)
2–1S(1)
1–0S(6)
1–0S(9)
1–0S(7)
3–2S(3)
0.003
2–1S(2)
1–0S(8)
2–1S(3)
0.002
0.001
0
1.7 1.8 1.9 2 2.1 2.2 2.3 2.4
Wavelength (microns)
Figure 5.17. IRAS 18151−1208. Top left: K-band continuum image of the IRS1 high mass
protostar; top middle: south-eastern structure pointing in the direction of the large scale jet;
top right: spectrum of IRS1; bottom: H2 ν1 − 0S(1) line maps and spectrum of a typical knot in
the jet (courtesy of Davis et al., 2004).
by cold (peak emission at 100 µm) and warm (peak emission at 20 µm) dust. Its bolo-
metric luminosity is ≈ 20,000 L⊙ , indicating that IRAS 18151–1208 is a high-mass star
in formation. Near-infrared images in the emission lines of molecular hydrogen H2 have
revealed two collimated jets in the IRAS 18151−1208 region, one of which is almost a
parsec in length. If these jets originate from the young stellar object, this would imply
that high-mass stars form via disc accretion. The K-band image also shows a bright cen-
tral peak (Figure 5.17 top left), denoted IRS1, roughly at the position of the mid-infrared
source and aligned with one of the jets. But is IRS1 really the high-mass star? And is
the jet really coming from that star? Unfortunately the error ellipse for the position
of the mid-infrared source is about 25′′ × 7′′ . The position of IRS1 in the near-infrared
colour-magnitude diagram supports the view that this star is indeed the high-mass star.
But this photometric identification is not conclusive, because emission from circumstel-
lar discs and dust extinction prevents the straightforward conversion from near-infrared
colours and magnitudes to temperature and luminosity.
The integral field spectroscopy from Davis et al. (2004) shows the H- and K-band
spectrum of IRS1 steeply rising at longer wavelengths (Figure 5.17 top right), indicating
a deeply embedded object. IRS1 also shows CO band-heads in emission, tracing dense
molecular gas (nH ≈ 1012 –1013 cm−3 and T ≈ 2000–5000 K) probably heated by UV pho-
tons impinging on the inner accretion disc surface. Alternatively, the CO emission could
come from the outer regions of accreting funnel flows, where the gas is heated by adiabatic
compression. The line maps from the [Fe II]λ1.644 µm and H2 emission (Figure 5.17 top
middle) show an ≈ 1′′ -long south-eastern structure extending from IRS1 in the direction
of one of the large-scale jets. Additional echelle spectroscopy supports the hypothesis that
this feature is indeed the base of the jet. The distant jet components show only H2 (Figure
2 ν =1−0S (1)
5.17 bottom right), but not [Fe II], indicating low excitation. The line ratio HH 2 ν =2−1S (1)
Science motivation for IFS and Galactic studies 191
is approximately 7–11, typical for shock heating (see excursion on H2 line diagnostics).
The analysis of the excitation diagram gives a temperature of ≈ 2000–2600 K. The knots
have a bow-like morphology (Figure 5.17 bottom left), and the kinematics are compa-
rable to what is seen in Herbig–Haro objects. In summary, Davis et al. (2004) find that
the knots in the IRAS 18151−1208 jets are similar to their counterparts from low-mass
protostars. The close association between the H2 features and the high-velocity CO emis-
sion observed at radio wavelengths suggests that the CO represents gas entrained by the
collimated jets. From the mass and momentum of the molecular gas and the luminosity
of the H2 features, it is clear that the flow must be powered by a massive source. To all
intents and purposes, the molecular jet appears to be a scaled-up version of jets from
low-mass young stellar objects.
Collectively, the observations add further support to the idea that massive stars are
formed through vigorous disc accretion, and that massive protostars drive collimated jets
in their earliest stages of evolution.
line ratio diagnostics in the optical). Vasconcelos et al. (2005) also provide an updated
mass loss rate for the 167–317 (LV2) proplyd from a simple model for the observed
Hα fluxes (see excursion): the mass loss for the photo-evaporated flow and the red-
shifted microjet are Ṁproplyd = (6.2 ± 0.6) × 10−7 M⊙ yr−1 and Ṁjet = (2.0 ± 0.7) ×
192 Frank Eisenhauer
200 400 600 800
50 50
40 40
30 30
20 20
10 10
0 0
a) Hα
Flux (× 10−15 erg cm−2 s−1)
Figure 5.18. 167–317 (LV2) proplyd in Orion. Top: Hα intensity and line profile map (left),
flux weighted rms (middle) and skewness (right); bottom: typical Hα line profile (left); data
minus simple Gaussian fit (right), revealing a blue- and a redshifted component (courtesy of
Vasconcelos et al., 2005).
10−8 M⊙ yr−1 , respectively. For a typical proplyd mass of ≈ 0.1 M⊙ , the mass loss rate
implies that the lifetime is less than ≈ 105 yr.
Scaled reflectance
0.89 0.87
Scaled reflectance 0.98 (0.17)
0.905 North Uranius Patera
(0.28) 1.10 0.92
Sirenum Fossae
(0.21)
0.975
0.91 Amazonis 2
1.10 (0.28) Terra Sirenum 1
0.93
(0.25)
0.86 Amazonis 1
(0.30)
0.925 1.00
0.885 Thersis/West Arsia
(0.32) Terra Cimmeris
1.00 1.0 0.98
(0.19)
8500 9000 9500 10,000 10,500 11,000 8500 9000 9500 10,000 10,500 11,000
Wavelength (Ångstroms) Wavelength (Ångstroms)
1.1
(normalized to 1.0 at 0.85 µm)
Scaled reflectance
1.05
Endmember B (0.29)
1.0
Endmember A (0.13)
0.95
c
8500 9000 9500 10,000
Wavelength (Ångstroms)
Figure 5.19. Spectra of the Marian surface. Left: typical bright regions characterized by the
absorption band from ferric iron oxides at ≈ 0.86 µm; right: typical faint regions characterized
by the absorption band from proxine at ≈ 0.93 µm and ≈ 0.96 µm; bottom: end-members used
for the linear decomposition of the spectra (courtesy of Martin et al., 1996).
2 2
PC
PC
2
1
α(X-ray), normalized
1
α(X-ray)
1.5 0
−1
1
−2
3 × 103 104 −2 −1 0 1 2
FWHM (Hb), km s−1 log FWHM(Hb), normalized
Figure 5.20. Geometric picture of principal component analysis (here illustrated for the X-ray
spectral index–Hβ line width relation for quasars). Left: measurements in natural units; right:
normalized units (mean subtracted, divided by standard deviation) and direction of principal
components (courtesy of Francis and Wills, 1999).
where j, k = 1, . . . , P . For our application, the principal components have the format of
a spectrum.
I/F
2.0
0.06
1.5
0.02 K-band
1.0
2.00 2.05 2.10 2.15 2.20 2.25 −80 −60 −40 −20 0 20
Latitude
Wavelength [µm]
Figure 5.21. Integral field spectroscopy of Titan. Top: spectral images in the K-band probing
the surface (left), troposphere (middle) and stratosphere (right); bottom left: typical K-band
spectra (dashed line shows the best fitting radiative transfer model); bottom right: tropospheric
haze enrichment (courtesy of Ádámkovics et al., 2006).
scattered as aerosol haze high in the lower stratosphere (40–200 km). At shorter wave-
lengths (e.g. 1.588 µm and 2.039 µm) – towards the limits of the H- and K-bands – we
see only weak absorption, and the observations trace the albedo of the surface. Interme-
diate wavelengths (e.g. 1.611 µm and 2.140 µm) probe accordingly intermediate altitudes,
and the spectral images show the tropospheric (0–40 km) haze distribution. The strato-
spheric haze has a higher concentration near the north (winter) pole. In contrast, the
tropospheric haze is enhanced near the south pole (Figure 5.21 bottom right). Surface
albedos are recovered at 60 mas (375 km) resolution in both the 1.5 µm and the 2.0 µm
windows, nicely tracing the high-resolution surface maps from the Cassini spacecraft.
For the detailed analysis the authors apply a two-stream, plane–parallel, radiative trans-
fer model to fit the observed spectra (see excursion). This modelling provides a tool to
determine the vertical haze profile and retrieve latitudinal trends in both stratospheric
and tropospheric aerosol extinction: the stratospheric aerosol extinction is measured to
increase at a rate of ≈ 0.65 ± 0.05% per degree latitude from 40S to 60N. In contrast, the
tropospheric haze is confined near the south pole at latitudes above 40S.
REFERENCES
Ádámkovics, M., de Pater, I., Hartung, M. ET AL. (2006), JGR, 111, 7
Allington-Smith, J., Murray, G., Content, R. ET AL. (2002), PASP, 114, 892
Bacon, R., Adam, G., Baranne, A. ET AL. (1995), A&AS, 113, 347
Bacon, R., Copin, Y., Monnet, G. ET AL. (2001a), MNRAS, 326, 23
Bacon, R., Emsellem, E., Combes, F. ET AL. (2001b), A&A, 371, 409
Bahcall, J.N., Tremaine, S. (1981), ApJ, 244, 805
Beck, T.L., Riera, A., Raga, A.C., Aspin, C. (2004), AJ, 127, 408
Binney, J., Tremaine S. (1987), Galactic dynamics, Princeton University Press
Bonnet, H., Abuter, R., Baker, A. ET AL. (2004), Msngr, 117, 17
Bower, R.G., Morris, S.L., Bacon, R. ET AL. (2004), MNRAS, 351, 63
Comte, G., Marcelin, M. (eds. 1995), ASPCS, 71
Davis, C.J., Varricatt, W.P., Todd, S.P., Ramsay Howat, S.K. (2004), A&A, 425, 981
Dougados, C., Cabrit, S., Lopez-Martin, L., Garcia, P., O’Brien, D. (2003), Ap&SS,
287, 135
Dumas, C. ET AL. in preparation
Eisenhauer, F., Schödel, R., Genzel, R. ET AL. (2003a), ApJ, 597, L121
Eisenhauer, F., Abuter, R., Bickert, K. ET AL. (2003b), Proc. SPIE, 4841, 1548
Eisenhauer, F., Genzel, R., Alexander, T. ET AL. (2005), ApJ, 628, 246
Emsellem, E., Cappellari, M., Peletier, R.F. ET AL. (2004), MNRAS, 352, 721
Förster Schreiber, N.M., Genzel, R., Lehnert, M.D. ET AL. (2006), ApJ, 645, 1062
Francis, P.J., Wills, B.J. (1999), ASPCS, 162, 363
Garcı́a, P.J.V. (2006), New Astronomy Review, 49, 590
Genzel, R., Thatte, N., Krabbe, A., Kroker, H., Tacconi-Garman, L.E. ET AL. (1996),
ApJ, 472, 153
Genzel, R. ET AL. (1997), in SINFONI a High Resolution Near-Infrared Imaging Spectrometer
for the VLT, www.mpe.mpg.de/SPIFFI/preprints/STC.210.ps.gz
Genzel. R., Pichon, C., Eckart, A., Gerhard, O.E., Ott, T. (2000), MNRAS, 317, 348
Genzel, R., Schödel, R., Ott, T. ET AL. (2003a), ApJ, 594, 812
Genzel, R., Schödel, R., Ott, T. ET AL. (2003b), Natur, 425, 934
Ghez, A.M., Duchêne, G., Matthews, K. ET AL. (2003), ApJ, 586, L127
Herbst, T.M., Beckwith, S.V.W., Glindemann, A. ET AL. (1996), AJ, 111, 2403
Kasper, M.E., Feldt, M., Herbst, T.M., Hippler, S., Ott, T., Tacconi-Garman, L.E.
(2002), ApJ, 568, 267
Kissler-Patig, M., Roth, M.M., Walsh, J.R. (2007), in Science Perspectives for 3D Spec-
troscopy, ESO Astrophysics Symposia, eds. Kissler-Patig, M., Wals, J.R., & Roth, M.M.
Springer
Krabbe, A., Genzel, R., Eckart, A., ET AL. (1995), ApJ, 447, L95
Lavalley, C., Cabrit, S., Dougados, C., Ferruit, P., Bacon, R. (1997), A&A, 327, 671
Lavalley-Fouquet, C., Cabrit, S., Dougados, C. ET AL. (2000), A&A, 356, L41
Leitherer, C., Schaerer, D., Goldader, J.D. ET AL. (1999), ApJS, 123, 3
Leonard, P.J.T, Merrit, D. (1989), ApJ, 339, 195
Levin, Y., Beloborodov, A.M. (2003), ApJ, 590, L33
Martin, P.D., Pinet, P.C., Bacon, R. ET AL. (1996), P&SS, 44, 859
McDermid, R., Bacon, R., Adam, G., Benn, C., Cappellari, M. (2004), Proc. SPIE, 5492,
822
McGregor, P.J., Dopita, M., Wood, P., Burton, M.G. (2001), PASA, 18, 41
Science motivation for IFS and Galactic studies 199
Monreal-Ibero, A., Roth, M.M., Schönberner, D., Steffen, M., Böhm, P. (2005), ApJ,
628, L139
Müller-Sánchez, F. ET AL. (2006), A&A, accepted, astro-ph/0604317
Nagar, N.M., Falcke, H., Wilson, A.S. (2005), A&A, 435, 521
Narayan (2002), in Lighthouses of the Universe: The Most Luminous Celestial Objects and
Their Use for Cosmology, Proceedings of the MPA/ESO Conference, 405
Osterbrock, D.E. (1989), Astrophysics of gaseous nebulae and active galactic nuclei, Univer-
sity Science Books
Reid, M.J., Brunthaler, A. (2004), ApJ, 616, 872
Rutten R.G.M., Benn, C.R., Méndez, J. (eds. 2006), New Astronomy Reviews, 49, 487
Schmidt, M. (1963), Natur, 197, 1040
Schödel, R., Ott, T., Genzel, R. ET AL. (2002), Natur, 419, 694
Shu, F.H., Adams, F.C., Lizano, S. (1987), ARA&A, 25, 23
Swinbank, A.M., Smail, I., Bower, R.G. ET AL. (2005), MNRAS, 359, 401
Tecza, M., Genzel, R., Tacconi, L.J. ET AL. (2000), ApJ, 537, 178
Van Breugel, W., Bland-Hawthorn, J. (2000), ASPCS, 195
Van Zyl, L., Charles, P.A., Arribas, S., Naylor, T., Mediavilla, E., Hellier, C. (2004),
MNRAS, 350, 649
Vasconcelos, M.J., Cerqueira, A.H., Plana, H., Raga, A.C., Morisset, C. (2005), AJ,
130, 1707
Weitzel, L., Krabbe, A., Kroker, H. ET AL. (1996), A&AS, 119, 531
6. Extragalactic studies and future integral
field spectroscopy science
LUIS COLINA
In this set of lectures, I review recent observational progress on extragalactic studies using
integral field spectroscopy (IFS) techniques, highlighting the importance of IFS for the
study of the nuclear regions of nearby galaxies, of low-z active galactic nuclei (AGN)
and massive star-forming galaxies, and of high-z galaxies, including lensed quasars,
lensing galaxies and bright submillimetre galaxies. Emphasis is given to the study of
(ultra)luminous infrared galaxies as examples of low-z systems where the physical pro-
cesses relevant to the formation and evolution of galaxies can be investigated in more
detail. Research projects involving future ground-based facilities and satellites are also
briefly presented.
6.1 Introduction
The use of IFS for extragalactic studies has burgeoned over the past 10 years and is
already becoming a standard observational technique used by several groups in many
different areas. Most IFS systems (INTEGRAL, GMOS, PMAS, SAURON, SINFONI,
VIMOS, etc.) allow us to simultaneously obtain spectra covering a wide spectral range
over a wide field of view (up to 1 arcmin square for VIMOS). These instruments in their
standard configurations provide low–intermediate spectral resolution (R of 1000 to 4000)
with a relatively low angular resolution (0.5′′ to 3.0′′ ). In addition, a few IFS systems, such
as OASIS on the William Herschel Telescope and SINFONI on the Very Large Telescope
(VLT), can provide very high angular resolution (i.e. 0.1′′ ) in the optical (OASIS) and
near-infrared (SINFONI) when combined with adaptive optics (AO) systems. Finally,
multi-object integral field spectrographs (GIRAFFE on the VLT) are already available
that are leading the way to future IFS systems currently in design (see Section 6.3). A
full account of the different available IFS systems and their characteristics can be found
in M. Bershady’s contribution to this Winter School.
Studies of galaxies at low and high redshift benefit greatly from the flexibility of IFS
systems. Detailed studies on different angular scales – from the nuclear, to circumnuclear
and external regions – can only be executed efficiently with IFS. Moreover, these systems
allow us to investigate simultaneously the two-dimensional structure of many astrophys-
ically important properties such as chemical composition, stellar populations, ionization
conditions, kinematics, etc. that are directly related to the formation and evolution of
galaxies.
Among the several areas of extragalactic research covered by IFS-based studies, this
contribution will focus on the study of (ultra)luminous infrared galaxies – (U)LIRGs –
explaining in detail the impact and use of IFS in the investigation of several topics, such
as the structure of the internal dust extinction, the properties of young stellar clusters, the
presence of different ionizing sources such as hidden AGN, (extra)nuclear star-forming
regions and shocks, and the kinematic properties of the nuclear and extended ionized
gas. All these aspects are common to many other research areas, such as the study of
the nuclear regions in nearby ellipticals and spirals, of the ionization structure in low-z
AGN, of extended nebulae in low- and high-z galaxies, and of studies of high-z galaxies,
including lensed objects. Due to space limitations, these studies will not be covered in
200
Extragalactic studies and future IFS science 201
10 arcsec
[O III] Hβ
Arcsec
SAURON WFPC2HST
E
Arcsec Arcsec Arcsec Arcsec
–100 +30 0 110
V [km/s] +100 0 σ [km/s] 170 –30 V [km/s] σ [km/s]
Figure 6.1. Maps of NGC 7742 taken with SAURON. The top panels show the emission line
intensity distributions of [O III] and Hβ, followed by a SAURON reconstructed image composed
of [O III], blue continuum and red continuum, and a similar image composed from HST/WFPC2
filters. The bottom row shows (from left to right) the derived gas velocity and velocity dispersion
fields, and the stellar velocity and velocity dispersion fields (figure reproduced from de Zeeuw
et al., 2002).
detail here and the reader is referred to the relevant references given in the following as
part of this introduction.
10 3.6 10 1000
∆d (arcsec)
∆d (arcsec)
0 3.3 0 900
10 3.8 10 1000
∆d (arcsec)
∆d (arcsec)
0 3.15 0 900
10 3000 10 1000
∆d (arcsec)
∆d (arcsec)
0 2000 0 900
Figure 6.2. Intensity (arbitrary units) and velocity (in km s−1 ) maps for NGC 5033 taken with
INTEGRAL. The intensity maps correspond to the stellar continuum (4700–5840 Å wavelength
interval), and the ionized gas traced by the [O III] and Hβ emission. The star and the circle
in the velocity maps represent the location of the intensity peak in the stellar continuum and
[O III] intensity maps, respectively (figure reproduced from Mediavilla et al., 2005a).
in this area. Detailed studies of the galaxy NGC 1068, considered as the prototype of
Seyfert 2 galaxies, include early work on the biconical ionization field (Arribas et al., 1996)
and on the kinematics of the central nuclear and circumnuclear region (Garcı́a-Lorenzo
et al., 1997, 1999). Similar individual studies of other well-known nearby galaxies on scales
of hundreds of parsecs and covering a wide range of activity from LINERs to Seyfert 1
galaxies have been undertaken from the early days of IFS to the present (see Figure 6.2
for an example; Arribas and Mediavilla, 1994; Arribas et al., 1997, 1999; Mediavilla et al.,
1997, 2005a; Garcı́a-Lorenzo et al., 2001; Ferruit et al., 2004). With the advent of near-IR
AO-based IFS systems such as SINFONI, the star formation in the inner regions close to
the AGN, the properties of the molecular gas in the central 100 pc region and its relation
to the torus is starting to be investigated on spatial scales of a few to several parsecs
(Davies et al., 2006; Mueller-Sánchez et al., 2006).
Not only has work on AGN covered radio-quiet AGN, but IFS has also been used
to investigate the kinematics and ionization of the interaction of the radio jet with the
surrounding interstellar medium in the radio galaxy 3C 120 (Sánchez et al., 2004; Garcı́a-
Lorenzo et al., 2005). VIMOS on the VLT is now also being used to obtain deep IFS of
the giant nebulae with sizes in excess of 100 kpc that surround high-z radio galaxies, and
Extragalactic studies and future IFS science 203
4 4
a)
2 2
∆d (arcsec)
0 0
–2 –2
–4 –4
4 2 0 –2 –4 4 2 0 –2 –4
4 4
b)
2 2
∆d (arcsec)
0 0
–2 –2
–4 –4
4 2 0 –2 –4 4 2 0 –2 –4
4 4
c)
2 2
∆d (arcsec)
0 0
–2 –2
–4 –4
4 2 0 –2 –4 4 2 0 –2 –4
∆a (arcsec) ∆a (arcsec)
Figure 6.3. Image representing six of the twenty continuum maps obtained for the lensed sys-
tem SBS 0909+532 using INTEGRAL. a) to c), left to right: Continuum central wavelengths are
8727 Å, 7233 Å, 5739 Å, 4447 Å, 4006 Å and 3487 Å. Changes in the relative brightness of the two
components as a function of wavelength are due to internal extinction within the intermediate
redshift lens galaxy (figure reproduced from Motta et al., 2002).
to establish their origin. The kinematics of the nebulae around the galaxy MRC2104-242
at a redshift of 2.49 is consistent with rotation, therefore implying a dynamical mass of
at least 3 × 1011 M⊙ in this system (Villar-Martı́n et al., 2006).
1 1
Arc seconds
Arc seconds
0 0
E E
–1 –1
N N
–1 0 1 –1 0 1
HST NICMOS H160 + Ho velocity field c) D d)
1 A 1
C
Arc seconds
Arc seconds
0 0
B
E E
–1 –1
0 kms–1 +270 N N
HST ACS I814 + GMOS IFU Lya
–1 0 1 –1 0 1
Arc seconds Arc seconds
Figure 6.4. a) Image of N2 850.4 combining HST ACS and NICMOS imaging. The image shows
a complex morphology, with at least three distinct components separated by 1 arcsec (8 kpc)
in projection. b) Infra-Red Telescope Facility (IRTF) Hα narrow-band image with the contours
from the NICMOS H-band image overlaid. This Hα narrow-band image shows a diffuse halo of
material distributed asymmetrically around the galaxy. c) The velocity field derived from UIST
IFS observations of the Hα emission line overlaid on the NICMOS H-band image. Components
A and B are separated by 50 ± 50 km s−1 whilst there is a velocity difference of 270 ± 50 km s−1
between components A and C. d) HST ACS I-band image with the Lyα intensity from the
GMOS IFS overlaid as contours (figures reproduced from Swinbank et al., 2005).
(Colina et al., 2001, 2005) into moderate-mass ellipticals through the process of merging
(Genzel et al., 2001; Tacconi et al., 2002; Dasyra et al., 2006) suggests a physical process
for forming early-type galaxies in the early Universe, such as those recently detected at
redshifts of 2–3 (Franx et al., 2003; Cimatti et al., 2004; Förster Schreiber et al., 2004;
van Dokkum et al., 2004).
Deep submillimetre and radio surveys (Chapman et al., 2003a, 2003b, and references
therein) are detecting a population of objects, the so-called sub-millimetre galaxies
(SMGs), that appear to be forming stars at rates of up to 1000 M⊙ yr−1 (i.e. up to
10 times more than ULIRGs), and that, according to their morphology (Chapman et al.,
2003b), luminosity and spectral energy distribution (Egami et al., 2004; Frayer et al.
2003), seem to be the high-z counterparts of the low-z luminous (LIRG) and ultralumi-
nous (ULIRG) infrared galaxies (for a review, see Blain et al., 2002). Mid-IR satellites
(Spitzer Space Telescope as well as millimetre interferometers, IRAM, ALMA) continue
to detect high-z LIRGs and ULIRGs in large quantities, as already shown by the data
taken with Spitzer (Charmandaris et al., 2004; Egami et al., 2004; Le Floc’h et al., 2004).
The majority of these galaxies at z < 1 are in the LIRG class, and they make a significant
contribution to the star formation density at 0.5 < z < 2 (Pérez-González et al., 2005).
Studies of the evolution of the morphological types of galaxies as a function of redshift
indicate that, as in low-z ULIRGs, mergers have played a major role in the shaping
of present-day galaxies. These studies (Conselice et al., 2003, 2005) conclude that the
fraction of galaxies consistent with undergoing a major merger increases with redshift
206 Luis Colina
HST/WFPC2 F814W
5 C'
C1
B21
0
D3
–5 B2
B1
B16
–5
Ha Hb
–5
5 0 –5 5 0 –5 5 0 –5
Figure 6.5. Continuum and emission line images for NGC 3690 obtained with INTEGRAL.
The contours represent the red stellar continuum at 6460 Å. The HST/WFPC2 I-band (F814W)
image is shown for comparison, with several regions of interest marked on it. Component B2
dominates the stellar continuum emission while components C1 and B1 are the strongest emis-
sion line sources (figure reproduced from Garcı́a-Marı́n et al., 2006).
for all galaxies, and that at redshifts above 2, more than 80% of the stellar mass is in
objects with peculiar morphology. There is also a predominance of irregular and complex
morphologies in high-z luminous SMGs, suggesting that major mergers are common in
these dusty star-forming galaxies (Chapman et al., 2003b). In addition, galaxies less
affected by dust – such as the Lyman break galaxies (LBGs) and UV-selected, high-z
star-forming galaxies – do often show irregular and distorted morphologies (Erb et al.,
2003, 2004; Giavalisco et al., 1996), indicating that interaction/merging is also involved
in these galaxies.
4 4 4
R2
Nn
2 2 2
0 0 0
–2 –2 –2
R1
Ns
–4 –4 –4
4 2 0 –2 –4 4 2 0 –2 –4 4 2 0 –2 –4
[OIII] Ha Hb
4 4 4
2 2 2
0 0 0
–2 –2 –2
–4 –4 –4
4 2 0 –2 –4 4 2 0 –2 –4 4 2 0 –2 –4
Figure 6.6. INTEGRAL images of the ionized gas and stellar light distribution in the central
regions of IRAS 12112+0305 as traced by different emission lines and by continuum windows
close to Hβ and Hα. The observed [O III] and Hβ emission is dominated by an extinction-free
extranuclear star-forming region. The HST I-band (F814W) image is shown for comparison
(figure reproduced from Colina et al., 2000).
examples). The wavelength dependence of the optical and near-IR stellar light distribu-
tion is mostly produced by the absorption due to the large amount of dust towards the
nucleus in these galaxies. A clear example of high nuclear extinction is shown in IRAS
12112+0305 where the red bright, compact southern nucleus (Ns ) appears as a weak
source in the Hβ-continuum light (Figure 6.6) seen through an extinction of about eight
magnitudes in the visual (Colina et al., 2000). In other galaxies, such as Arp 220 and
IRAS 17208−0014, the presence of a strong central dust lane obscures the true nucleus in
the optical. As a result, the optical nucleus, identified as the high surface brightness peak,
does not spatially coincide with the near-IR nucleus but is about 1 kpc apart in these
galaxies (Arribas et al., 2001; Arribas and Colina, 2003). In several ULIRGs investigated,
the internal extinction produces a wavelength dependence of the relative brightness of the
heavily obscured nuclear regions with respect to the less obscured external van Dokkum
regions of the host galaxy while keeping the overall stellar light distribution, even in
interacting pairs.
The observed ionized gas and stellar distributions show decoupled structures with
different morphologies in many (U)LIRGs. These differences are due to the presence of
extranuclear gas ionized by a central active galactic nucleus (Mrk 273; Colina et al., 1999;
Arp 299; Garcı́a-Marı́n et al., 2006), the presence of bright, tidally induced extranuclear
star-forming regions along the tidal tails (IRAS 14348−1447; Monreal-Ibero, 2004) or
far away (9 kpc) from the nucleus of the galaxy (IRAS 12112+0305; Colina et al., 2000)
and the presence of circumnuclear gas ionized by a nuclear, dust-enshrouded starburst
(Arp 220; Arribas et al., 2001). In addition, the complex, non-axisymmetric structure
of the internal extinction in these galaxies produces additional wavelength-dependent
effects such as changes in the relative brightness of the nuclear regions with respect
208 Luis Colina
to less absorbed star-forming regions (IRAS 08572+3915; Arribas et al., 2000), or to
extranuclear regions located several kpc away from the nucleus (IRAS 12112+0305).
The ionized gas distribution traced by the emission lines shows a different structure
in some of the galaxies. Since these emission lines trace different ionization mechanisms
(shocks, young stars, AGN), their two-dimensional light structure represents the spatial
distribution of these different ionizing sources within a given galaxy. This is clearly seen
in the two nearest galaxies studied so far, Arp 299 and Arp 220. The peaks of the [O I]
λ6300 Å, [O III] λ5007 Å, Hα and [S II] λλ6717,6731 Å emitting regions in Arp 299 are not
spatially coincident and present substructure even on scales of 200 pc or less (Figure 6.5
and Garcı́a-Marı́n et al., 2006). The high surface brightness plumes in Arp 220, detected
up to 4 kpc from the nucleus, present a complex ionization structure as identified by the
changes in the [N II]/Hα emission line ratio measured along the plumes (Colina et al.,
2004).
a) b)
5 C'
C1
B21
0
D3
–5 B2
B1
B16 HST F190N HST F160W
1.50
c) d)
5 1.25
1.01
E(B–V)
0 0.77
0.53
–5 0.29
0.05
5 0 –5 5 0 –5
Figure 6.7. a) Panel showing the galaxy NGC 3690 the HST/NICMOS continuum-subtracted
Paα emission; b) the HST/NICMOS H-band continuum; c) the INTEGRAL Hα extinction
corrected emission and d) the extinction spatial distribution obtained with the Hα/Hβ line
ratio. The contours represent the red stellar continuum at 6460 Å as obtained with INTEGRAL
(figure taken from Garcı́a-Marı́n et al., 2006).
Seyfert 2 (strong high excitation [O III] λ5007 Å emission line) nuclei. In most ULIRGs,
however, the presence of a dust-enshrouded AGN is not evident in the optical, based
only on the nuclear spectrum. Integral field spectroscopy allows us to investigate the
two-dimensional ionization structure of the nuclear and circumnuclear gas searching for
traces of optically dust-enshrouded AGN. These types of AGN can be traced by the pres-
ence of extended extranuclear high-excitation ionized regions similar in size and structure
to those detected in nearby AGN. Some ULIRGs, such as Mrk 273 (Colina et al., 1999)
and IRAS 11087+5351 (Garcı́a-Marı́n et al., in preparation), show the presence of highly
ionized [O III] emitting regions at distances of several kiloparsecs from the nucleus of the
galaxies. Some other galaxies show a well-defined ionizing cone morphology, as expected
from the standard AGN model (Urry and Padovani, 2000 and references therein). One
of these cones, detected in the galaxy NGC 3690, is in the process of merging with IC
694 and forms the luminous infrared system Arp 299 (Figure 6.8; Garcı́a-Marı́n et al.,
2006). According to the standard diagnostic diagrams, three different types of ionization,
HII-LINER and AGN-like, are identified. The line ratio maps clearly show the presence of
an extranuclear Seyfert-like nebula with a well-defined conical morphology, with its apex
in region B1, and located at a projected distance of 1.5 kpc from its apex. This highly
ionized conical structure in NGC 3690 is also detected at larger distances from B1 in the
210 Luis Colina
Figure 6.8. Diagnostic diagram and line ratio maps derived for the LIRG Arp 299 (IC 694
+ NGC 3690) using INTEGRAL with the SB3 bundle. a) [O III] λ5007 Å/Hβ versus [O I]
λ6300 Å/Hα diagnostic diagram with all the data displayed. The symbol code is as follows:
triangles represent values within the ionization cone, and at distances of less than 2 kpc from
region B1; stars represent the region within the cone at distances of 2–4 kpc from B1; crosses
mark the interface region between the two galaxies that form the Arp 299 system (IC 694 and
NGC 3690); and filled circles represent the regions in the two galaxies (defined as the regions
within the isocontours). b)–c) [OIII] λ5007 Å/Hβ and [O I] λ6300 Å/Hα maps, respectively. The
nuclei of galaxies IC 694 and NGC 3690 are identified with regions A and B1, respectively. The
white line delineates the ionization cone with its apex in B1 (figure reproduced from Garcı́a-
Marı́n et al., 2006).
[O I] λ6300 Å/Hα map (Figure 6.8). However, the ionization associated with the outer
regions of the cone is LINER-like rather than Seyfert, indicating that within the cone
there are two ionization regimes well separated spatially. Therefore, the two-dimensional
ionization maps strongly support photoionization by a radiation cone escaping from a
central dust-enshrouded AGN source located in B1. This AGN has also been recently
identified at other wavelengths (Della Ceca et al., 2002; Ballo et al., 2004; Gallais et al.,
2004).
Extragalactic studies and future IFS science 211
Figure 6.9. Spatial distribution of the spectra (centred on the Hα-[N II]) in the circumnuclear
region of Arp 220 obtained using the bundle SB2 of INTEGRAL (figure reproduced from Arribas
et al., 2001).
Figure 6.10. Some examples showing the complex spectra in the nuclear regions of Arp 220
and the fitted components indicating the presence of spatially separated velocity components.
Numbers represent the fibre as displayed in the previous figure (reproduced from Arribas et al.,
2001).
Figure 6.11. Velocity field and spatial distribution of the three kinematically distinct gaseous
components in the circumnuclear regions of Arp 220. These velocity components are associated
with rotational motions (A) and radial flows (figure reproduced from Arribas et al., 2001).
velocities imply a dynamical mass (Mdyn ) of 3.5 × 1010 M⊙ within a radius of 1.5 kpc.
This relatively high value indicates a large mass concentration in the nuclear region of
Arp 220, as already inferred by the presence of 5 × 109 M⊙ of molecular gas in a nuclear,
0.5 kpc size disc (Scoville et al., 1997).
A broad velocity component (B in Figure 6.11) with an average width (FWHM) of
815 km s−1 is identified over an extended region of 1 kiloparsec in diameter, at about
600 pc northwest of the dust-enshrouded nucleus, and blueshifted by 300 km s−1 with
respect to the system velocity. A third, high-velocity component (C in Figure 6.11) with
peak-to-peak velocities of 1000 km s−1 is detected mostly southeast of the nucleus and
orientated along the kinematic minor axis of the rotating component (for details, see
Arribas et al., 2001). The two-dimensional distribution and kinematics of components
B and C are consistent with a bipolar gas outflow characterized by an opening angle
Extragalactic studies and future IFS science 213
of about 90 degrees, and orientated perpendicularly to the nuclear rotating disc of gas
traced by component A. These overall gas kinematics agree with the basic predictions
of the starburst-driven galactic wind scenario proposed by Heckman et al. (1990) for
galaxies such as (U)LIRGs with powerful nuclear starbursts.
Velocity fields
The velocity fields of the ionized gas in the ULIRGs investigated with IFS show a complex
structure (see Figure 6.14), inconsistent in general with large, kiloparsec-scale, ordered
motions indicating that rotational motions do not dominate the velocity field. Only one
galaxy (IRAS 17208−0014) shows a velocity field consistent with that of a rotating
214 Luis Colina
Figure 6.12. a) Spatial distribution of the spectra around the redshifted Hα–[N II] spectral
region over the entire 75′′ × 40′′ extended nebulae in Arp 220. Spectra taken with INTEGRAL
using the low angular resolution bundle (SB3 with 2.′′ 7 diameter fibres) and three different point-
ings. b) Image of the stellar light distribution in the central regions and faint extended envelope
of Arp 220 as traced by a line-free narrow-band red continuum. c) Image of the ionized gas
distribution in Arp 220 as traced by the [N II] λ6484 Å emission line. The bright circumnu-
clear region, the elongated plumes and the extended diffuse lobes are easily identified. Note the
clear differences in the ionized gas distribution with respect to the stellar light presented above
(figures reproduced from Colina et al., 2004).
Extragalactic studies and future IFS science 215
Integral
5800
10 (a) Arp 220
Velocity field
5600
0
∆ d (arcsec)
kms–1
–10 5400
–20 5200
–30 5000
40 20 0 –20
∆ a (arcsec)
Integral
5800
10 (b) Arp 220
Velocity field
5600
0
∆ d (arcsec)
kms–1
–10 5400
–20 5200
–30 5000
40 20 0 –20
∆ a (arcsec)
Figure 6.13. Velocity field of the extended ionized gas in Arp 220 obtained from the [N II]
λ6584 Åemission line. The contours superimposed represent the [N II] emission (a), and the
stellar light distribution, respectively (b). Orientation as in Figure 6.12 (figure reproduced from
Colina et al., 2004).
disc of about 3 kpc in size (Arribas and Colina, 2003), while one of the three velocity
components identified in the central regions of Arp 220 (see Section 6.2.5, Circum-nuclear
regions and Arribas et al., 2001) could also trace a nuclear rotating disc, already detected
in molecular gas (Downes and Solomon, 1998). In addition to IRAS 17208−0014 and
Arp 220, there are other galaxies (e.g. IRAS 08572+3915) in which the major photometric
and kinematic axes lie close in projection and could be considered as candidates for
rotating systems (Arribas et al., 2000). For the rest of the sample (seven galaxies in
six ULIRGs), the largest velocity variations, characterized by peak-to-peak differences
of 200–400 km s−1 , do not show a convincing case for rotation on scales of a few to
several kpc. Higher angular resolution AO-based IFS is needed to elucidate the presence
of rotation in the inner regions (≤ 1–2 kpc) of ULIRGs in general. Evidence for inward
flows (IRAS 15206+3342; Arribas and Colina, 2002), outflows associated with nuclear
starbursts (Arp 220; Arribas et al., 2001), flat velocity fields (Mrk 273, IRAS 15250+3609)
and motions probably associated with tidal tails (IRAS 12112+0305, IRAS 14348−1447;
Monreal-Ibero, 2004) and with the merger process (Arp 220; see Section 6.2.5.2) are also
present in these systems. The sample is still small but the results indicate that tidally
induced motions dominate the velocity field of the extended ionized nebula on kiloparsec
scales. However, there is so far no clear trend between the observed velocity field and the
phase of the interaction/merging process traced by the stellar structure, i.e. intermediate
phases identified with pairs (i.e. IRAS 12112+0305, IRAS 14348−1447), more advanced
216 Luis Colina
MRK 273
5 500 450
375 225
kms–1
0 250 0
125 –225
–5
0 –450
IRAS 05572+3915
5 150 200
113 100
kms–1
0 75 0
38 –100
–5 0 –200
Arp 220
5 400 300
300 150
kms–1
0 200 0
100 –150
–5
0 –300
IRAS 15208+4318
5 150 200
113 100
kms–1
0 75 0
38 –100
–5
0 –200
IRAS 17208–0144
5 200 250
150 125
kms–1
0 100 0
50 –125
–5 0 –250
IRAS 15250+3808
5 300 200
225 100
kms–1
0 150 0
75 –100
–5 0 –200
IRAS 14348–1447
5 300 300
225 150
kms–1
0 150 0
75 –150
–5 –300
0
IRAS 12112+0306
5 300 300
225 150
kms–1
0 150 0
75 –150
–5
0 –300
5 0 –5 5 0 –5 5 0 –5 5 0 –5 5 0 –5
Figure 6.14. Panel presenting the IFS-based kinematics of a sample of low-z ULIRGs in dif-
ferent evolutionary states. The stellar and ionized gas distribution are traced by the optical
continuum (second column from the left), and Hα line (centre column). The kinematic proper-
ties of the ionized gas are represented by the velocity dispersion map (second column from the
right) and the velocity field (right column). The Hα contours are superposed on the velocity
dispersion and velocity field maps. The near-IR (NICMOS F160W filter) HST imaging is also
shown on the left side for comparison (figure reproduced from Colina et al., 2005).
Extragalactic studies and future IFS science 217
phases represented by double nuclei (i.e. Arp 220, Mrk 273) and evolved single-nucleus
galaxies (i.e. IRAS 17208−0014, IRAS 15250+3609).
Analysis of a larger sample of ULIRGs, also including the lower luminosity LIRGs, is
needed in order to characterize the gas kinematics and to establish the role and impor-
tance of starburst-driven winds, tidal-induced flows and ordered rotational motions as
the interaction/merger process evolves.
Figure 6.15. [N II] λ6584 Å/Hα versus [S II] λλ6717,6731Å/Hα ratios. This figure can be
divided into several regions. In the left bottom corner is located the region occupied by typical
H II regions while the top right corner of the plot is the locus for a typical LINER-like spectrum.
Different symbols represent values for different galaxies. Values for the fibres associated with the
circumnuclear region are indicated with filled symbols, while those for the other fibres appear
with open symbols. Models of Dopita and Sutherland (1995) for shocks have been superposed.
Those without precursor are indicated with solid lines, while those with precursor are plotted
using dashed lines. At the beginning of each line the magnetic parameter Bn−1 / 2 (µG cm3 / 2 )
is shown. Shocks velocity range from 150 to 500 km s−1 for models without precursor and from
200 to 500 for models with precursor. The long-dashed line indicates the [N II] λ6584 Å/Hα
versus [S II] λλ6717,6731 Å/Hα values predicted by photoionization by a power-law model for
a dusty cloud at ne = 103 cm−3 and Z = Z⊙ (Groves et al., 2004). The dotted line indicates the
locus for an instantaneous burst model of 4 Myr, Z = Z⊙ , IMF power-law slope of −2.35 and
Mu p = 100 M⊙ (Barth and Shields, 2000); dust effects have been included and ne = 103 cm−3
(figure reproduced from Monreal-Ibero et al., 2006).
in these regions. Further evidence for the dominant role of shocks in the ionization of these
regions comes from the positive correlation measured between the velocity dispersion and
ionization status of the ionized gas in the extranuclear regions for several galaxies. Figure
6.16 represents such a correlation using the [O I] λ6300 Å/Hα line ratio, i.e. the most
reliable diagnostic ratio to detect ionization by shocks (Dopita and Sutherland, 1995),
and where each data point represents the value for a specific spectrum (fibre), i.e. differ-
ent position in the extranuclear nebula but excluding the circumnuclear region, obtained
with IFS. The fact that the extranuclear gas in a sample of galaxies covering different
dynamical phases of the merging process follows a well-defined relation between the line
ratio, and the velocity dispersion reinforces the scenario that tidal-induced shocks are the
dominant ionization source on large scales (> 2–3 kpc). However, the fact that the ioniza-
tion and velocity dispersion of the ionized gas varies from galaxy to galaxy indicates that
Extragalactic studies and future IFS science 219
Figure 6.16. Relation between the shock-tracer [O I] λ6300 Å/Hα line ratio and the velocity
dispersion for the extended, extranuclear gas regions in ULIRGs. Different symbols represent
different galaxies. The dashed horizontal line marks the frontiers between the H II region and
LINER-type ionization (figure reproduced from Monreal-Ibero et al., 2006).
6.5 Acknowledgements
Based on observations with the William Herschel Telescope operated on the island of La
Palma by the ING in the Spanish Observatorio del Roque de los Muchachos of the Insti-
tuto de Astrofı́sica de Canarias. Also based on observations with the NASA/ESA Hubble
Space Telescope, obtained at the Space Telescope Science Institute, which is operated by
the Association of Universities for Research in Astronomy, Inc. under NASA contract
number NAS5-26555.
Much of the work presented here has been supported by the Ministry for Education and
Science of Spain under its National Programme for Astronomy and Space Science. I would
like to thank my collaborators over the years: Santiago Arribas, Ana Monreal-Ibero,
Macarena Garcı́a-Marı́n, and Almudena Alonso-Herrero. Finally, a lot of the material
presented during the lectures was kindly made available by many colleagues who allowed
me to use material given in their PowerPoint presentations. My thanks to Drs Cardiel,
Davies, Falcón-Barroso, Hammer, Jarvis, Kuntschner, Lemoine-Busserole, McDermid,
Mollá, Swinbank, and Villar-Martı́n.
REFERENCES
Arribas, S., Colina, L. (2002), ApJ, 573, 576
Arribas, S., Colina, L. (2003), ApJ, 591, 791
Arribas, S., Mediavilla, E. (1994), ApJ, 437, 149–161
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B. (1996), ApJ, 463, 509
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B., del Burgo, C. (1997), ApJ, 490, 227
Arribas, S., Mediavilla, E., del Burgo, C., Garcı́a-Lorenzo, B. (1999), ApJ, 511, 680
Arribas, S., Colina, L., Borne, K.D. (2000), ApJ, 545, 228
Arribas, S., Colina, L., Clements, D. (2001), ApJ, 560, 160
Arribas, S., Bushouse, H., Lucas, R.A., Colina, L., Borne, K.D. (2004), AJ, 127, 2522
Arribas, S., Ferrint, P., Jakobsen, P. (2007), in Science Perspectives for 3D Spectroscopy.
ESO Astrophysics Symposia, eds. M. Kissler-Patig, M. Roth & J. Walsh. Springer
Extragalactic studies and future IFS science 223
Bacon, R., Copin, Y., Monnet, G. ET AL. (2001), MNRAS, 326, 23
Bacon, R., Bauer, S.-M., Bower, R. ET AL. (2004), in Ground-based Instrumentation for
Astronomy, eds. A.F.M. Moorwood & I. Masanori, Proceedings of the SPIE, Vol. 5492,
1145
Baldwin, J.A., Phillips, M.M., Terlevich, R. (1981), PASP, 93, 5
Ballo, L., Braito, V., Della Ceca, R. ET AL. (2004), ApJ, 600, 634
Barth, A.J., Shields, J.C. (2000), PASP, 112, 753
Blain, A. W., Smail, I., Ivison, R.J., Kneib, J.-P., Frayer, D.T. (2002), Physics Reports,
369, 111
Bushouse, H.A., Borne, K.D., Colina, L. ET AL. (2002), ApJS, 138, 1
Chapman, S.C., Blain, A.W., Ivison, R.J., Smail, I.R. (2003a), Natur, 422, 695
Chapman, S.C., Windhorst, R., Odewahn, S., Yan, H., Conselice, C. (2003b), ApJ, 599,
92
Charmandaris, V., Uchida, K.I., Weedman, D. ET AL. (2004), ApJS, 154, 142
Cimatti, A., Daddi, E., Renzini, A. ET AL. (2004), Natur, 430, 184
Colina, L., Arribas, S., Borne, K.D. (1999), ApJ, 527, L13
Colina, L., Arribas, S., Borne, K.D., Monreal, A. (2000), ApJ, 533, L9
Colina, L., Borne, K., Bushouse, H. ET AL. (2001), ApJ, 563, 546
Colina, L., Arribas, S., Clements, D. (2004), ApJ, 602, 181
Colina, L., Arribas, S., Monreal-Ibero, A. (2005), ApJ, 621, 725
Conselice, C.J., Bershady, M.A., Dickinson, M., Papovich, C. (2003), AJ, 126, 1183
Conselice, C.J., Blackburne, J.A., Papovich, C. (2005), ApJ, 620, 564
Dasyra, K. M., Tacconi, L.J., Davies, R.I. ET AL. (2006), ApJ, 651, 835
Davies, R. I., Thomas, J., Genzel, R. ET AL. (2006), ApJ, 646, 754
Della Ceca, R., Ballo, L., Tavecchio, F. ET AL. (2002), ApJ, 581, L9
de Zeeuw, P.T., Bureau, M., Emsellem, E. ET AL. (2002), MNRAS, 329, 513
Dı́az-Santos, T., Alonso-Herrero, A., Colina, L., Ryder, S.D., Knapen, J.H. (2007),
ApJ, 661, 149
Dopita, M.A., Sutherland, R.S. (1995), ApJ, 455, 468
Downes, D., Solomon, P.M. (1998), ApJ, 507, 615
Egami, E., Dole, H., Huang, J.-S. ET AL. (2004), ApJS, 154, 130
Emsellem, E., Cappellari, M., Peletier, R.F. ET AL. (2004), MNRAS, 352, 721
Erb, D.K., Shapley, A.E., Steidel, C.C. ET AL. (2003), ApJ, 591, 101
Erb, D. K., Steidel, C.C., Shapley, A.E., Pettini, M., Adelberger, K.L. (2004), ApJ,
612, 122
Farrah, D., Rowan-Robinson, M., Oliver, S. ET AL. (2001), MNRAS, 326, 1333
Ferruit, P., Mundell, C.G., Nagar, N.M. ET AL. (2004), MNRAS, 352, 1180
Flores, H., Hammer, F., Puech, M., Amram, P., Balkowski, C. (2006), A&A, 455, 107
Förster Schreiber, N.M., van Dokkum, P.G., Franx, M. ET AL. (2004), ApJ, 616, 40
Förster Schreiber, N.M., Genzel, R., Lehnert, M.D. ET AL. (2006), ApJ, 645, 1062
Franx, M., Labbé, I., Rudnick, G. ET AL. (2003), ApJ, 587, L79
Frayer, D.T., Armus, L., Scoville, N.Z. ET AL. (2003), AJ, 126, 73
Frayer, D.T., Reddy, N.A., Armus, L. ET AL. (2004), AJ, 127, 728
Gallais, P., Charmandaris, V., Le Floc’h, E. ET AL. (2004), A&A, 414, 845
Ganda, K., Falcón-Barroso, J., Peletier, R.F. ET AL. (2006), MNRAS, 367, 46
Garcı́a-Lorenzo, B., Mediavilla, E., Arribas, S., del Burgo, C. (1997), ApJ, 483, L99
Garcı́a-Lorenzo, B., Mediavilla, E., Arribas, S. (1999), ApJ, 518, 190
Garcı́a-Lorenzo, B., Arribas, S., Mediavilla, E. (2001), A&A, 378, 787
Garcı́a-Lorenzo, B., Sánchez, S.F., Mediavilla, E., González-Serrano, J.I., Chris-
tensen, L. (2005), ApJ, 621, 146
Garcı́a-Marı́n, M., Colina, L., Arribas, S., Alonso-Herrero, A., Mediavilla, E. (2006),
ApJ, 650, 850
Garcı́a-Marı́n, M. ET AL., in preparation
Gardner, J.P., Mather, J.C., Clampin, M. ET AL. (2006), SSRv, 123, 485
Genzel, R., Tacconi, L.J., Rigopoulou, D., Lutz, D., Tecza, M. (2001), ApJ, 563, 527
Giavalisco, M., Steidel, C.C., Macchetto, D. (1996), ApJ, 470, 189
Groves, B.A., Dopita, M.A., Sutherland, R.S. (2004), ApJS, 153, 9
Heckman, T.M., Armus, L., Miley, G.K. (1990), ApJS, 74, 833
Le Floc’h, E., Pérez-González, P.G., Rieke, G.H. ET AL. (2004), ApJS, 154, 170
224 Luis Colina
Lutz, D., Spoon, H.W.W., Rigopoulou, D., Moorwood, A.F.M., Genzel, R. (1998), ApJ,
505, L103
McDermid, R.M., Bacon, R., Kuntschner, H. ET AL. (2006), New Astronomy Review, 49,
521
McDowell, J.C., Clements, D.L., Lamb, S.A. ET AL. (2003), ApJ, 591, 154
Mediavilla, E., Arribas, S., Garcı́a-Lorenzo, B., del Burgo, C. (1997), ApJ, 488, 682
Mediavilla, E., Arribas, S., del Burgo, C. ET AL. (1998), ApJ, 503, L27
Mediavilla, E., Guijarro, A., Castillo-Morales, A. ET AL. (2005a), A&A, 433, 79
Mediavilla, E., Muñoz, J.A., Kochanek, C.S. ET AL. (2005b), ApJ, 619, 749
Monreal-Ibero, A. (2004), Estudio cinemático y de fuentes de ionización de galaxias infrar-
rojas ultraluminosas con espectroscopa bidimensional, PhD, Universidad de La Laguna
Monreal-Ibero, A., Arribas, S., Colina, L. (2006), ApJ, 637, 138
Motta, V., Mediavilla, E., Muñoz, J.A. ET AL. (2002), ApJ, 574, 719
Motta, V., Mediavilla, E., Muñoz, J.A., Falco, E. (2004), ApJ, 613, 86
Mueller Sánchez, F., Davies, R.I., Eisenhauer, F. ET AL. (2006), A&A, 454, 481
Pérez-González, P.G., Rieke, G.H., Egami, E. ET AL. (2005), ApJ, 630, 82
Rieke, G.H., Low, F.J. (1972), ApJ, 176, L95
Sánchez, S.F., Garcı́a-Lorenzo, B., Mediavilla, E., González-Serrano, J.I., Chris-
tensen, L. (2004), ApJ, 615, 156
Sarzi, M., Falcón-Barroso, J., Davies, R.L. ET AL. (2006), MNRAS, 366, 1151
Scoville, N.Z., Yun, M.S., Bryant, P. (1997), ApJ, 484, 702
Sharples, R.M., Bender, R., Lehnert, M.D. ET AL. (2004), in Ground-based Instrumenta-
tion for Astronomy, eds. A.F.M. Moorwood & I. Masanori, Proceedings of the SPIE, 5492,
1179
Soifer, B. T., Sanders, D.B., Madore, B.F. ET AL. (1987), ApJ, 320, 238
Solomon, P.M., Downes, D., Radford, S.J.E., Barrett, J.W. (1997), ApJ, 478, 144
Swinbank, A.M., Smith, J., Bower, R.G. ET AL. (2003), ApJ, 598, 162
Swinbank, A.M., Smail, I., Bower, R.G. ET AL. (2005), MNRAS, 359, 401
Swinbank, A.M., Bower, R.G., Smith, G.P. ET AL. (2006a), MNRAS, 368, 1631
Swinbank, A.M., Chapman, S.C., Smail, I. ET AL. (2006b), MNRAS, 371, 465
Swinyard, B.M., Rieke, G.H., Ressler, M. ET AL. (2004), in Optical, Infrared and Millimeter
Space Telescopes, eds. J.C. Mather, Proceedings of the SPIE, Vol. 5487, 785
Tacconi, L.J., Genzel, R., Lutz, D. ET AL. (2002), ApJ, 580, 73
Tecza, M., Baker, A.J., Davies, R.I. ET AL. (2004), ApJ, 605, L109
Tully, R.B., Fisher, J.R. (1977), A&A, 54, 661
Urry, M., Padovani, P. (2000), PASP, 112, 1516
van Dokkum, P.G., Franx, M., Förster Schreiber, N.M. ET AL. (2004), ApJ, 611, 703
Veilleux, S., Kim, D.-C., Sanders, D.B. (1992), ApJS, 522, 113
Villar-Martı́n, M., Sánchez, S.F., De Breuck, C. ET AL. (2006), MNRAS, 366, L1
Wright, G.S., Rieke, G.H., Colina, L. ET AL. (2004), in Optical, Infrared and Millimeter
Space Telescopes, eds. J.C. Mather, Proceedings of the SPIE, 5487, 653
7. Tutorials: How to handle 3D
spectroscopy data
SEBASTIÁN F. SANCHEZ, BEGOÑA GARCÍA-LORENZO AND
ARLETTE PÉCONTAL-ROUSSET
7.1 Introduction
Integral field spectroscopy (IFS) is a technique to obtain both spatial (x,y) and spec-
tral (λ) information of a more or less continuous area of the sky simultaneously on the
detector. Only a few instrumental concepts allow 3D information on 2D detectors to be
obtained, and all of these are based on field splitters such as fibre bundles, lens array,
or image slicers (see Figure 7.1) to sample the field of view. Each sampled element is
then dispersed using a classic spectrograph and produces a spectrum on the detector.
Depending on the field splitter used, the geometry of the spectra on the detectors may be
very different. This diversity leads to the creation of very specific reduction techniques
and/or packages, i.e. one per instrument built (e.g. P3d, Becker, 2001). Combined with
the inherent complexity of 3D techniques, such software diversity has reduced the use of
IFS for decades to a handful of specialists, mainly those involved in the teams building
such instruments.
Conscious that this would be a handicap IFS specialists Walsh and Roth (2002) have
started to standardize techniques and tools for integral field units (IFU). Recently, the
Euro3D Research Training Network (RTN), whose aim was to promote 3D spectroscopy
all over Europe (Walsh and Roth, 2002), made a great effort to create a standard data for-
mat (Kissler-Patig et al., 2004) for storing and exchanging 3D data, developing an applica-
tion programming interface, API (Pécontal-Rousset et al., 2004), to ease the use of such a
data format and creating a visualization tool (Sánchez, 2004) usable by any existing IFU.
All these tools are freely distributed to the community (http://www.aip.de/Euro3D).
However, there is still a lack of software packages of general use for most of the data
reduction and data analysis processes (reduction steps, line fitting, gas and stellar
kinematics, etc.) of 3D data. The astronomical community faces a plethora of differ-
ent pipelines and tools specific to each instrument. We have developed a set of tools and
procedures useful for most of the IFS data; these will be explained in these tutorials,
as they were presented in the Instituto de Astrofı́sica de Canarias (IAC) Winter School.
Many of these tools are still in the development process, and we are distributing the
current version without any guarantee that they will work in the optimal way. The aim
is to provide simple and standardized tools, and also to teach the user the main logical
steps he or she will face in trying to reduce and then analyse his or her 3D data.
! Section 7.2 describes the most widely used data formats to store 3D spectroscopy
The layout of the chapter will be as follows:
! Section 7.3 describes some of the existing tools to analyse 3D data before the Euro3D
data.
! Section 7.4 describes the installation requirements of the different packages used in
network.
! Section 7.5 introduces E3D, the Euro3D visualization tool (Sánchez, 2004).
these tutorials.
225
226 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
Figure 7.1. Layout of the main techniques used to perform integral field spectroscopy
(Allington-Smith and Content, 1998).
! Section 7.10 describes the data reduction process with fibre-feed spectrographs.
refraction.
! To display an RSS image with DS9, first select the File menu and then select Open.
systems.
! To display datacubes (i.e. 3D images) with DS9, do the same as above (i.e. select the
File menu and then the Open option). Once you have choosen the file to be loaded,
a pop-up window comes up entitled ‘Data Cube’, with a scroll bar that allows you to
scroll through spectral pixels (i.e. wavelength). Moving that scroll bar, DS9 displays
! To display datacube stores in multi extension FITS file, use the File menu and then
the slice corresponding to the selected wavelength (monochromatic image).
the ‘Open Other’ option and the ‘Open Multi Ex as Data Cube’ option.
DS9 can be also run from the shell, entering ds9 followed by the name of the FITS
image (e.g. ds9 test.fits). For multiple extension FITS files, use the metacube option of
DS9 (e.g. ds9 -medatacube test.fits).
Figure 7.2. DS9 displaying an RSS image. The different rows correspond to the various spectra
of a particular area of an object (an active galaxy in this case). The horizontal axis is wavelength.
Vertical lines covering all the spectra are sky lines. A few emission lines from the observed object
can also be identified in this frame.
n being the value of the third dimension to display. If it is a multiple extension format,
use
display test.cube1.fits[n][*,*],
where n is the extension number to display.
You can also inspect the different spectra using the splot command:
splot test.cube.fits[10,10,*].
The RSS images (2D images) may also be visualized using the display command:
display test.ms.fits.
Tutorials: How to handle 3D spectroscopy data 229
Figure 7.3. Visualizing 3D data using IRAF and imarec. The snapshot of the console shows
the different IRAF tasks used to visualize RSS data.
1
INTEGRAL (http://www.iac.es/proyect/integral/) is a 3D instrument at the William
Herschel Telescope (WHT) (Arribas et al., 1998).
230 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
7.3.3 Starlink
The STARLINK project (http://starlink.rl.ac.uk/) compiled (until 2005) a set of interac-
tive data reduction and analysis tools for astronomical data. For example, in the STAR-
LINK software package list we can find FIGARO, a great reduction package for reducing
spectroscopic data, which also includes image and datacube manipulation capabilities
(for more information see http://starlink.rl.ac.uk/star/docs/sun86.htx/sun86.html). By
default, FIGARO accesses data files in STARLINK’s NDF (Extensible N-Dimensional
Data Format) data format, but it can also access other data formats, like FIGARO’s old
DST (Data Storage Technology) format, FITS and IRAF formats. By the way, IRAF is
a software facility of the STARLINK project. Check the following web page for a com-
plete list of STARLINK software: http://starlink.rl.ac.uk/static www/soft2 LSF.html.
Unfortunately, the STARLINK project has already finished and software packages such
as FIGARO are unsupported.
As an example of the STARLINK software capabilities for analysing 3D data, we will
use DIPSO, a friendly spectrum analysis program, for fitting lines. Once the user has
started the STARLINK and DIPSO sessions, the plotting device should be defined using
the command dev xwindow. Then the alasrd command reads an ASCII file containing
the spectral data and relevant information. The initial values of the fitting parameters
(e.g. central wavelengths, full-width-at-half-maximum or FWHM, etc. of the different
line components) are defined using the task elfinp. Finally, the elfopt command runs the
fit. Figure 7.4 shows both the complete sequence of DIPSO commands for fitting lines on
a spectrum and a plot of the results. Unfortunately, for 3D data this procedure should
be done once per spectrum and the number of spectra may be huge. Moreover, after the
fitting, maps of the line parameters (such as intensity, and FWHM) are highly desirable,
but neither DIPSO nor any other STARLINK tasks can do these. Other software tools
must be used to create such maps.
Figure 7.4. Line fitting using DIPSO (from the STARLINK software).
graphical user interface (GUI), they provide the user with all the necessary data reduction
steps, plus a set of generic graphical tools to view or plot data, assign keywords, etc. They
also provide analysis tools like image reconstruction (maps), spectra summation over an
aperture, line fitting (developed externally and currently being incorporated), merging
of datacubes, etc. They all handle different data formats, and are currently debugged to
fully support Euro3D. Binaries of the XOasis software are made available to the user com-
munity and may be downloaded from the Oasis Webserver, while others may be obtained
through the consortia who develop the instruments. The last ones usually contain
more sophisticated analysis tools like line-of-sight velocity distribution (LOSVD) fitting,
etc.
Another example is IDA. INTEGRAL team developed IDA, the Integral Data Anal-
ysis tool (Garcı́a-Lorenzo et al., 2002). IDA was written in interactive data language
(IDL) and offers a widget-based interface. IDA uses an interactive display tool, ATV
(for more information about ATV, see http://www.physics.uci.edu/∼barth/atv/), which
allows control of images similar to other display tools like DS9. IDA plots individual
spectra, creates and displays maps, overplots spectra on recovered images, zooming in
and zooming out, and calculates velocity fields through the cross-correlation technique.
Although IDA was written to work with INTEGRAL data, it can be also used for similar
data from other IFSs through minor software changes. Figure 7.6 shows some of the IDA
capabilities for analysing INTEGRAL data.
232 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
7.3.5 Exercises
In this section, we present a few exercises to visualize 3D data using traditional astro-
! datacube → NAME.cube.fits;
! Euro3D format data → NAME.e3d;
! Position Table → NAME.pt.
Using IRAF
To run IRAF, type cl or ncl at any xterm or xgterm:
cl> imred NAME RSS.fits
cl> specred NAME RSS.fits
It is possible to zoom in, zoom out, fit lines, plot different lines, with all the options of
the gtools. For more help, see help/splot.txt and help/gtools.txt
Figure 7.6. The INTEGRAL Data Analysis (IDA) tool capabilities for visualizing and
analysing 3D data.
! FIT3D, a fitting tool for 3D spectroscopy (Sánchez et al., 2004; Sánchez et al., in
! R3D, a software package for reducing fibre-based IFS (Sánchez et al., in preparation).
preparation);
They all have a C-core written using the I/O Lyon-C Library and its development envi-
ronment (Pécontal-Rousset et al., 2004). The PGPLOT graphical library (Pearson, 1995)
is used also for plotting purposes and, in some cases, Tcl/Tk libraries for GUIs. In many
cases, perl is also used as a scripting language, or as a prototyping environment.
! Astro::FITS::CFITSIO.
and the following perl modules (http://www.cpan.org).
! Math::Approx.
! Math::Derivative.
! Math::FFT.
! Math::Matrix.
! Math::Spline.
! Math::Stat.
! PDL.
! PGPLOT.
! Statistics::OLS.
We recommend the user installs the source code of all these packages. You can down-
load this from the following web page: http://www.iac.es/winters/2005/software3D.html.
When available, pre-compiled binaries are also included for both Linux or Mac OSX, to
ease the installation for users who are not very familiar with software building, and who
are using standard Linux/Mac OSX distributions.
2
The new gcc compiler does not include anymore g77, so you must install it separately.
Tutorials: How to handle 3D spectroscopy data 235
aip.de/Euro3D/). A copy of version 1.2 is included there (E3D io LCL-1.2.tar.gz). To
built it, just follow these instructions:
(i) First create a directory in your computer, that we name hereafter as yourDIR. Copy
the compressed tar file into this directory yourDIR.
(ii) Uncompress and untar the file.
(iii) Go to the created directory, e.g. cd E3D io LCL-1.2
(iv) Run the configure program: ./configure
(v) Run the make program: make
(vi) Define the environment variables: IFU PATH and IFU DEFAULT FMT; e.g. for a tcsh
shell:
setenv IFU PATH yourDIR3 ;
setenv IFU DEFAULT FMT Euro3D;
(vii) To check the library, go to the directory:
cd checklib/exec
and run the test program:
./check E3D io
If you do not have any error messages, then the install is OK.
3
The IFU PATH should refer to yourDIR and not to yourDIR/E3D io LCL-1.2.
236 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
(iv) Go to the scripts directory (i.e. cd scripts) and execute the install script
(./install.pl). This script will customize the tk e3d.tcl according to your direc-
tory tree.
(v) Add the path to the directory user/bin (under the E3D directory) to your environ-
ment variable PATH.
(vi) Test your E3D installation by executing the script tk e3d.tcl, and then loading
the file test.e3d that you will find under the scripts directory.
level functions, including routines calling the Euro3D I/O library, routines to access
the shared memory server (SHM), and basic functions needed to plot and analyse
the data. These functions may be invoked by any external program, which needs to
! The tk e3d Tcl/Tk interpreter. This is a stand-alone program embedding the Tcl/Tk
read/write Euro3D data, or uses the SHM.
interpreter with Euro3D routines. These routines wrap the previous C-functions in
Tcl, and can be used, within the tcl interpreter, to load/save Euro3D files, plot
single or co-added spectra, plot monochromatic/polychromatic maps, interpolate
! A number of stand-alone tools, which may be used from the command line, may be
these, save maps in FITS format, etc.
helpful in accessing Euro3D files. Perhaps the most interesting routine is any2Euro3D,
converting single-group IFS data from any instrument into the Euro3D Format.
Together with this C-core, we provide a Tcl/Tk GUI, which calls the Euro3D-Tcl routines
(tk e3d.tcl). The GUI is organized around three different windows corresponding to
! The Stacked Spectra Inspector, or main window, includes the main menu with various
three different ways of investigating the data:
options to handle the input and output data and to pop up the other windows. Once
loaded, the data are automatically plotted in its canvas, using the stacked-spectra
representation. Moreover, a command-line prompt has been included in this window
238 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
Figure 7.7. E3D Stacked Spectra Inspector: this is the E3D main GUI window, from which
all the others can be created. It displays all the spectra contained in the 3D file stacked row by
row.
Figure 7.8. E3D Spaxels Inspector: this is the GUI provided for plotting monochromatic or
polychromatic maps. It is possible to select different spaxels, to be displayed subsequently on
the Spectral Inspector.
to directly call any Euro3D-‘tcl flavor’ routines. Figure 7.7 gives a snapshot of this
! The Spaxels Inspector provides the user with a main canvas for viewing
window.
Figure 7.9. E3D Spectral Inspector: this is the GUI provided for plotting spectra.
! The Spectral Inspector has a canvas for viewing/plotting the spectra corresponding
to the spaxels selected from any of the two previous windows. It also includes a
menu with different options to handle the spectra. Figure 7.9 gives a snapshot of
this window.
7.5.1 Working with E3D: the GUI
We describe here, with a few examples, the capabilities of E3D. As we quoted above, E3D
can handle data in the Euro3D format, its default format, or use the RSS and datacube
formats. All of these are converted internally to the same structures, so, once loaded,
– In the top menu of the main window, select Load in the File menu. A pop-up
window appears.
– Change from default fits extension to e3d.
– In the top menu of the main window select Import RSS in the File menu. A
pop-up window appears.
– Select the NAME RSS.fits RSS file.
! Loading datacubes:
– Select the NAME RSS.pt position table.
– In the top menu of the main window select Import cube in the File menu. A
pop-up window appears.
– Select the file NAME.cube.fits.
You can now follow the exercises. In order to be able to run the exercises, first change to
the directory tut1 and open the E3D GUI running the tk e3d.tcl command.
Exercise 1: Try to load your own data. First identify what format your data are
in (read the reduction tool notes if necessary).
If you need to build a position table, you have to know that the position tables in E3D
are ASCII files with four columns, where the first row gives the geometry of the dataset
and the remaining rows contain the positions, as below:
TYPE SIZE1 SIZE2 TYPE ID
INDEXi Xi Yi TYPE IDi
...
240 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
TYPE is the shape of the spaxels and can be: R for a rectangle shape, S for a square, C
for a circle, and H for a hexagon. Here is an example of such a table:
H 0.2 0.2 1
1 0.0 0.0 1
2 0.5 0.0 1
3 0.5 0.5 1
...
7.5.2 Quick look over the various windows of the E3D GUI
As quoted before, the GUI of E3D has three windows corresponding to various inspections
of the data:
! In the top-level menu of the main window, select Import cube in the File menu. A
By default, the main window displays only 256 spectra (i.e. not the total number of
spectra) as a 2D image, where each row represents a single spectrum. E3D automatically
sets the cuts of intensities for the look-up table, determining the median and the standard
deviation of the intensities plotted. Sometimes, the estimate is not good enough and it is
better to set the values manually. To change these maximum and minimum values, just
fill the Max and Min entries in the left-hand menu. Note: A range of [−1200] would be
acceptable to visualize the data. For a better contrast it is sometimes useful to tune the
Brightness and Contrast settings too, edit the corresponding entries, or just move the
corresponding scroll bars.
Exercise 2: Change the maximum, minimum, brightness and contrast settings until you
get a proper visualization of your data.
To browse the spectra (in cases where you have more than 256 spectra) you can move the
scroll bar in the bottom part of the left-hand menu, below the label Range of Spectra
plotted. It is also possible to change the number of plotted spectra, editing the values
of the two entries labelled First (the first spectrum plotted) and Range (the number of
spectra plotted).
It is possible to zoom over a given spectral region, by selecting this region and clicking
twice over the main display with a left click of the mouse. The selected wavelength range
appears in the W1 and W2 entries in the left-hand menu. You can also edit these entries
directly, if you want a more accurate selection of the wavelength bounds. In order to
return to the default range click on the Release button.
Tutorials: How to handle 3D spectroscopy data 241
E3D always uses the same method to select elements. The left button of the mouse is
used to select a window and the right button for sending this selection to another window
(inspector). We will experiment with this later on.
Exercise 4: Play with the zooming capabilities until you get used of them.
The combination of zooming capability and range selection is useful to look at specific
wavelength and spatial ranges; e.g. for emission line objects.
Exercise 5: Try to locate some of the emission lines in this object.
Exercise 9: Practise using the spectra inspector. Change the maximum and minimum,
the brightness and the contrast in order to visualize the selected spectra properly.
As in the main window, it is possible to zoom over spectral regions and/or select them,
using the same settings as above. You can also display the average slice of a selected
number of spectra by right-clicking on the mouse.
Exercise 14: Experiment with the various spaxel selection methods. Select spectra one
by one, switch to aperture selection, and to the pseudo-slit selection. See the differences
in the spectra inspector and experiment until you get used to the different methods.
! tools for determining the shift and angle of the image along the wavelength;
tools can be divided into two groups:
Exercise 15: Experiment with DAR correction (dir: tut3). Run E3D and
import the cube star no dar.cube.fits. Open the spaxels inspector and, using the
movie effect, check the DAR effect. Click on the peak correction on the main window
menu bar (DAR tools) and perform the correction. Store the result in another file (e.g.
star dar.cube.e3d). Note that the correction is not optimal.
! plot spectra: This has the same behaviour as right-clicking on the mouse.
option, but you may also:
! plot profile: This plots the intensity profile of the slice currently displayed along
the selected spaxels. It is particularly useful when combined with the pseudo-slit
selection.
246 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
Exercise 19: Hit the s key on top of the spaxels inspector display. Look at the infor-
mation labelled displayed on top of it. It should say Pseudo-Slit Selection. Select a
cut that goes through the centre of the image, and select the plot profile option. The
intensity profile along the pseudo slit will appear in the spectra inspector, fitted with a
Gaussian.
! create object: An object stands for a collection of selected spaxels. Once selected,
a new entry will appear on the Objects list box in the left-hand menu. A maximum
of nine different objects can be created; if this number is exceeded, the objects are
overwritten. Objects are the best way to store your previous selections. To get back
to any object, just click on it in the Objects list box. The spectra corresponding
to the selection of spaxels stored in this object will be displayed in the spectra
inspector.
Exercise 20: Play with the object creation (dir: tut4). On the spaxels inspec-
tor select the central spaxels and store the selection as an object. Clear the selection
and repeat the process with the outer part of the central square, and with the external
circle of spaxels. Open the spectral inspector and select alternatively each of the objects.
Look at the different selections of spectra displayed.
It is possible to store a selected group of spaxels (or object) in a separate Euro3D file,
using the save object option. Just click on one of your objects (or select a group of
spaxels) and select this menu option. If you want to see what the result looks like, you
! reverse object: this option is used to reverse the spaxels selection stored in an
can load it in the main window.
object (i.e. select the ones that were not selected, and vice versa). Click this option
and then on one of your objects.
Exercise 21: Separate sky and objects fibres (dir: tut4). Import the INTEGRAL
SB2.rss.fits file, and display a slice in a given wavelength range using the spaxels
inspector. One by one select all the spaxels in the outer ring, i.e. those corresponding to
the sky fibres. Once selected, store the selection as an object twice. Reverse the selection
of one of these objects, so that you have the central square selected. Apply the two avail-
able selections and store the results with the save object option in two different files:
obj.e3d and sky.e3d. Load these files in E3D to see the result.
Exercise 22: Mask regions creating objects (dir: tut3). Import the VIMOS.
cube.fits file in E3D. Open the spaxels inspector. Select a given wavelength range and
display the corresponding slice on the spaxels inspector. You may see four bad regions
corresponding to dead fibres. Select them and store the selection as an object. Reverse
the selection and store the object as good data.e3d. Load the newly created Euro3D file
and display the result in the spaxels inspector. You have removed the bad data.
! Blink: Allows blinking between slices stored in the eight-length small display buffer.
This gathers the basic tools that operate over the slices:
Exercise 23: Play with the blinking facility (dir: tut4). Import the INTEGRAL
SB2.rss.fits file and display different slices at various wavelength ranges (hint: select
continuum dominated regions and emission line regions). Then, click on the blink button
and select, by left-clicking the mouse on the small displays, the slices you want to blink
Tutorials: How to handle 3D spectroscopy data 247
in between. Click again on the blink button to start the blinking, and once again when
you want to stop it.
! Arithmetic: This allows the user to perform simple arithmetic computations
between two buffered slices. It brings up a small pop-up window whose parame-
ters are: the two slices to be processed (using their buffer number) and the number
of the buffer in which to store the result. You select the type of the computation by
pulling down the Arithm menu at the bottom of the window.
Exercise 24: Map of emission lines, continuum subtracted (dir: tut4). Load
the obj.fits (created before), and display two slices each corresponding to the wave-
length region of [OIII] and of a continuum region nearby. Subtract the continuum to the
[OIII] slice. This will be a raw estimation of the continuum-free emission-line map.
Keyboard shortcuts
! The ‘s’ key switches from the pseudo-slit selection mode to the spaxel-to-spaxel (or
There are a few keys that may be used for different operations:
! The ‘z’ and ‘u’ keys allow the user to zoom over the image. Hit the z key, with the
aperture), selection.
mouse cursor located on the upper corner of the area to zoom and once more with the
mouse cursor over the lower corner. A zoom over the selected region will be displayed
in the spaxels inspector. To unzoom, hit the u key over the spaxels inspector to go
back to the display of the whole image.
Exercise 25: Play with the zooming capability (dir: tut3). Import the VIMOS.
cube.fits file in E3D. Open the spaxels inspector. Select a given wavelength range and
display the corresponding slice on the spaxels inspector. Zoom in and out over different
regions until you get used of it.
Exercise 26: Testing the saving capabilities of the spectra inspector (dir:
tut4). Import the INTEGRAL SB2.rss.fits file and display a slice using the spaxels
inspector. Select different groups of spaxels and display the corresponding spectra in the
spectra inspector. Save the selected spectra as a Postscript and a GIF file (hard-
copies), then as a 2D RSS image, a FITS table and an ASCII table. Visualize the outputs
using external tools, such as ghostview, xview, iraf or DS9. Repeat the operation until you
get used to it.
248 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
The Spectra menu
This menu handles the storage in memory (or buffering) of average spectra. The capa-
bility is similar to the object creation in the spaxels inspector (see above). These
spectra may be plotted, one-by-one or simultaneously, or may be used to perform basic
computations, like those exhibited in the spaxels inspector using slices.
Clicking on Save Spectrum stores a new spectrum in the memory (up to a maximum
of nine), and a new entry appears in the list box labelled Spectra. In order to display
any of the buffered spectra, just click on the corresponding entry in the list box.
It is possible to view several spectra on the same display by selecting the tag labelled
single/multiple plot on the pull-down menu, and then selecting the spectra to be
plotted using the spectra list box. Each spectrum will be plotted using a different
colour in order to distinguish them. Click on the Clear Multiple selection option to
clear the screen and create a new selection.
Exercise 27: Test the spectra buffering capability (dir: tut4). Import the
INTEGRAL SB2.rss.fits file and display a slice using the spaxels inspector. Select dif-
ferent groups of spaxels and plot the corresponding spectra in the spectra inspector.
Store each of these spectra in a separate slot. Select each of the spectra to view them in
the spectra inspector display. Change from single to multiple selection and repeat the
selection to see the difference. Store them as Postscripts and/or GIF files, and visualize
them using external tools.
spectra. A window pops up (similar to the slice arithmetic one), requesting the
selection of two spectra (from the list box) and the type of arithmetic operation to
perform (pull-down menu Arithm on the bottom of the window). The result will be
stored in the first available slot.4
Exercise 28: Get used to the spectra arithmetic capability (dir: tut4).
Import the INTEGRAL SB2.rss.fits file and display a slice using the spaxels inspector.
Select various groups of spaxels and plot the corresponding spectra in the spectra
inspector. Store each of these spectra in a separate slot. Select spectra to view them
in the display of the spectra inspector. Perform different arithmetic operations with
the selected spectra and visualize them using the single or multiple display options to
compare the results.
! Sky Subtraction allows to subtract one of the buffered spectra from the dataset.
This option is mostly used to perform a sky subtraction. To use it, just select this
option in the menu, and then select the spectrum to be subtracted (i.e. the sky
spectrum).
Exercise 29: Sky subtraction (dir: tut4). Import the INTEGRAL SB2.rss.fits file
and display a slice using the spaxels inspector. Select the spaxels in the outer ring (the
sky-fibres ones), and visualize the average sky spectrum in the spectra inspector. Store
this spectrum in one of the memory slots, and perform a sky subtraction using this spec-
trum as the sky spectrum. Once the subtraction is performed it is possible to save the
whole dataset on disk. Go back to the main window and save it as an RSS or a Euro3D
dataset. It is also possible to remove the sky spaxels by storing the sky spaxels selection as
4
There is a known bug when trying to use spectra from the first position of the slot.
Tutorials: How to handle 3D spectroscopy data 249
an object, performing a reverse object operation, and then storing the object (e.g. as
sobj.fits for RSS format or sobj.e3d for Euro3D datacube), as described in the Exer-
cise above.
! Fit Single Emission line: This fits the plotted spectrum with a single emission
are provided:
line (a Gaussian function), plus a continuum (as a threshold). The four free param-
eters of the fitting are the intensity of the Gaussian, the FWHM, the central wave-
length and a constant for the continuum. The result of the fit will be overplot in
red (in contrast to the data which are plotted in white). The output of the fitting
routine is printed in the xterm.
Exercise 30: Fit an emission line (dir: tut4). Import the sobj.fits or sobj.e3d
file and select the slice whose wavelength range is centred on the [OIII] emission line, with-
out any other emission line pollution, using the spaxels inspector. Plot the corresponding
spectrum using the spectral inspector. Fit this spectrum using the Fit Single Emission
line.
! Full Automatic Kinematics Analysis: This performs a kinematics analysis using
a single emission-line fitting (as described above) per spectrum. Being an automatic
procedure, the result is not fully reliable but must be considered as a quick estimation
of the kinematics on the object. It fits each spectra in the dataset with a single
Gaussian function (the emission line) plus a constant (the continuum). Once the
fitting is done, it creates four maps using the output of the fit, i.e. a map of the
line intensity, a map of the FWHM, a map of the central wavelength and a map of
the continuum emission. The fit is performed over the currently selected wavelength
range.
Exercise 31: Full kinematics estimate (dir: tut4). Import the sobj.fits or
sobj.e3d file and select a slice centred on the [OIII] emission line, without pollution
of any other emission line, using the spaxels inspector. Plot the corresponding spectrum
using the spectral inspector. Fit this spectrum using the Fit Single Emission line.
Once this first fit is performed, click on the Full Automatic Kinematics Analysis.
The routine loops over the spectra of the dataset. Once finished, four new maps will be
created in the spaxels inspector, one for each free parameter of the model (see above).
Check visually the different maps.
300
OH Meinel bands
200
Intensity fv (µJyarcsec–2)
V
100
OI OE
B 5577
OI
6302 8645
U
NaD
OI
5890/6
Hg 6364
Hg NI? 5460
4858 O lines 5199
s
0
4000 5000 6000 7000 8000 9000
Wavelength (Ångstroms)
Figure 7.10. Typical sky spectrum in the optical wavelength range obtained on a moonless
night at the Observatorio del Roque de los Muchachos (ORM) on La Palma island (Spain).
Most of the emission features identified in the sky spectrum are due to airglow. The Meinel
rotation-vibration bands of [OH] dominate the sky background spectrum redward of 6500 Å.
The typical broad band filters for astronomical observations are marked (figure from Benn and
Ellison, 1998).
distinct [OH] lines can be identified which are spectrally unresolved at a resolution of
∼ 17,000. The problem of subtracting such number of sky lines is a difficult one because
the absolute and relative intensities of [OH] lines are strongly temporal and spatially
dependent. For this reason, the accuracy of the sky subtraction is much better in the
continuum than in atmospheric emission lines, and the large number of [OH] lines makes
an accurate sky background subtraction in the red difficult.
But the accuracy of the sky background also depends on instrumental features. Typi-
cal instrumental problems such as flexures in the spectrograph, aberrations in the optics
or variations in the spectral efficiency of the detector also cause bad features when sub-
tracting the sky spectrum.
Therefore, atmospheric and instrumental factors complicate such an a priori easy task
of subtracting the sky spectrum from the object spectra. Figure 7.11 illustrates troubles
found when subtracting the sky background on 3D data.
Depending on the science to be done, sometimes it is very important to subtract the
sky background with a high level of accuracy. In some cases – for example in infrared
observations – the object may have a surface brightness much fainter than the sky (for
example to study faint objects). Because IFS is usually used for observing extended
objects, most of the time it is necessary to sample the sky background far enough from the
target object (sky background exposures) using a telescope off-set. The problem is worse
when IFS is used with adaptive optics, when the point-spread function (PSF) delivered
by the telescope and the adaptive optics system should be the same in both target and
background exposures for a proper and accurate sky subtraction (for a description of
Tutorials: How to handle 3D spectroscopy data 251
Figure 7.11. The DS9 displays viewing the same RSS image before (left) and after (right) the
sky background subtraction. On the left display, several emission lines are present. After the
sky background subtraction (on the right), some residuals remain where the brightest sky lines
were located.
different ways of removing the sky background, see Allington-Smith and Content, 1998).
Therefore, the accuracy of the sky background subtraction often depends on observing
strategies.
MIDAS task for removing the sky spectrum from long-slit data. This program makes
252 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
a polynomial fit of the sky spectrum using two different windows above and below
the object spectrum. The final sky spectrum to be subtracted is obtained from a
combination of the spectra of these two windows.
The sky background component can be also estimated using the commands
DEFINE/SKY and EXTRACT/SKY within MIDAS. The command DEFINE/SKY
allows the user to define the off-set limits of up to two sky windows, usually on
both sides of the object spectrum. The command EXTRACT/ECHELLE performs
an average extraction of the sky background using these off-sets. The two extracted
subtract the sky from long-slit data by polynomial fitting in the spatial direction
of two defined regions on either side of an object of interest. FIGARO users can
also use the command SCNSKY, which uses the minimum median algorithm to
create a sky spectrum from an image without any lines clear of stars. FIGARO also
provides the SKYLINER task removing a sky spectrum normalized by the height of
the 5577 Å [OI] emission line.
– two integers, the first one always left to 0 (to identify the file), and a second giving
the number of individual functions included in the model;
– two floats, the first one being the goal value for the reduced χ2 and then the goal
value for the variation of this value between two consecutive iterations. When one
5
This is a three-column ASCII file containing: ID, WAVELENGTH and FLUX.
Tutorials: How to handle 3D spectroscopy data 253
– an integer flag indicating if this parameter is to be fitted (= 1) or if it will stay
fixed (= 0);
– the lower bound for the parameter;
– the upper bound for the parameter;
– an integer flag indicating if this parameter is free (= −1) or if it is linked to the
same parameter of another function of the model (= to the number of the function
! The linking method allows the fitting of lines that share some properties (e.g. lines
in the configuration file to which it is linked, 1 being the first one).
that are kinematically coupled) or the inclusion of lines for which the line ratio is
known (e.g. the line ratio between [OIII] 5007 Å and [OIII] 4959 Å). There are two
ways to link parameters:
– additive: the parameter of this function is equal to the parameter of the function
indicated by the element number (5) plus a constant. In this case element number
(3) indicates the constant to be added and element number (4) must be set to 0.
– multiplicative: the parameter of this function is equal to the parameter of the
function indicated by the element number (5) multiplied by a constant. In this
case element number (3) indicates the multiplicative factor, and element number
(4) must be set to 1.
Although many functions are described with fewer than nine parameters, the nine lines
are mandatory (but possibly filled with zeroes). The currently available functions in
! gauss1d one-dimensional Gaussian, whose parameters are (in the following order):
FIT3D, to build a model for Observatoire de Lyon, are:
(1) the central wavelength, (2) the peak intensity and (3) the σ of the Gaussian
! eline a Gaussian emission line, whose parameters are (in the following order): (1)
function (describing the width, FWHM = 2.345 σ).
the central wavelength, (2) the integrated flux, (3) the σ of the Gaussian function
! poly1d a n-order polynomial function (n < 9), whose parameters are the coefficients
(FWHM = 2.345 σ) and (4) the velocity in km s−1 .
! back a spectrum background, included in the list of files BACKGROUND LIST (one file
of the polynomial function.
per line), whose parameters are: (1) order of the background file in the list, (2) the
intensity scaling factor, (3) the σ of a Gaussian function to convolve the background
spectra to match the resolutions and/or the velocity dispersions, and (4) the velocity
in km s−1 to shift the spectrum.
! The lines and the background are kinematically coupled, defining a system (see
The constraints included are as follows:
As shown in this example, one of the powerful features of this tool is the possibility of
defining systems. A system is a set of emission lines which share the same velocity and
velocity dispersion. The dispersion is included as a dispersion in terms of wavelength and
not of velocity, which may be a problem for large wavelength ranges. This co-moving
set of emission lines may be fundamental for deblending several components in complex
systems. The continuum can also be coupled kinematically (or not) to a system.
For example:
254 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
0 4 0.2 0.001
...
eline
poly1d
5007 1 0 0 -1
0.01 1 -10 10 -1
1.0 1 0.2 20 -1
0 0 0 0 -1
4.6 1 0 0 -1
0 0 0 0 -1
12000 1 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
eline
back
4959 1 0 0 -1
1 0 0 0 -1
0.33 1 0.33 1 1
10.0 1 0.2 200 -1
4.6 1 0 0 1
4.6 1 0 0 1
12000 1 0 0 1
12000 1 0 0 1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
...
The result of the fitting process is displayed on the screen and stored in the following
! out.fit spectra, an ASCII file including a header line with the number of fitted
output files:
functions (i.e. the number of function included in the model), followed by a line per
function. Each line contains 19 values:
– the label of the model, as described above;
– the values of its nine parameters;
! out config.fit spectra, a configuration file where the output values of the fit are
– the nine corresponding error estimates.
used as first-guess values. This file may be useful when fitting iteratively similar
! out mod res.fit spectra, a five-column ASCII table. The columns are:
spectra with the same model.
– the wavelength;
– the original flux at this wavelength;
– the derived model;
– the residual; and
– the residual continuum.
The syntax is as follows:
fit spectra SPECTRUM.TXT CONFIG FILE [BACKGROUND LIST NBACK]
fit spec back.pl , a perl wrapper to create configuration files
This program creates the configuration file needed by fit spec back. It requires an
input spectrum in the ASCII format (see above), an ASCII file listing the emission lines
to be fitted (see tut4/emission lines.txt), the wavelength range to be fitted, a file
containing the list of backgrounds to use (if not applicable, type none), the number of
backgrounds files, and a file defining the wavelength bounds of the spectral regions to
mask.
Tutorials: How to handle 3D spectroscopy data 255
The emission lines file is an ASCII file with two columns, one for the wavelength of the
emission line, and the other one for the name (or label) of the line; for example:
5006.84 [OIII]5007
4958.91 [OIII]4959
4861.32 Hbeta
The file defining the spectral regions to mask is a two-column ASCII file with respectively
the lower bound (minimum wavelength) and the upper bound (maximum wavelength) of
the region to mask. The syntax is:
fit spec back.pl SPECTRUM.TXT LINE FILE START WAVELENGTH
END WAVELENGTH BACK LIST NBACK MASK LIST [CONFIG]
Exercise 32: Fitting and deblending emission lines (dir: tut4). Using E3D,
import the sobj.fits or sobj.e3d file and display a slice with a wavelength range
centred in the [OIII] emission line, avoiding other emission lines. Plot the corresponding
spectrum. Save this spectrum as an ASCII table (e.g. as spec cen.txt). Exit from E3D.
Enter on the shell command-line:
fit spec back.pl spec cen.txt emission lines.txt 4800 5300 none 0 none
Then answer the following questions:
Number of systems to include:2
Redshift of the system number 0:0.03
Systemic Velocity=8993.77374 km/s
Flexibility of the velocity (km/s):500
GAUSSIAN sigma of the line (GAUSS,min,max): 3 0.1 40
Redshift of the system number 1:0.03
Systemic Velocity=8993.77374 km/s
Flexibility of the velocity (km/s):500
GAUSSIAN sigma of the line (GAUSS,min,max): 40 0.1 1000
HeII 4686.0 at 4826.58 found
Include? (y/n):n
[OIII]5007 5006.84 at 5157.0452 found
Include? (y/n):y
Flux (Flux, min, Max)?:2000 100 5000
Included
[OIII]4959 4958.91 at 5107.6773 found
Include? (y/n):y
Flux (Flux, min, Max)?:2000 100 5000
Included
Hbeta 4861.32 at 5007.1596 found
Include? (y/n):y
Flux (Flux, min, Max)?:2000 100 500
Included
[FeII] 4889.62 at 5036.3086 found
Include? (y/n):n
[FeII] 4905.34 at 5052.5002 found
Include? (y/n):n
[FeII] 5111.6299 at 5264.978797 found
256 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
Include? (y/n):n
System 0 contains 3 lines
HeII 4686.0 at 4826.58 found
Include? (y/n):n
[OIII]5007 5006.84 at 5157.0452 found
Include? (y/n):n
[OIII]4959 4958.91 at 5107.6773 found
Include? (y/n):n
Hbeta 4861.32 at 5007.1596 found
Include? (y/n):y
Flux (Flux, min, Max)?:2000 100 5000
Included
[FeII] 4889.62 at 5036.3086 found
Include? (y/n):n
[FeII] 4905.34 at 5052.5002 found
Include? (y/n):n
[FeII] 5111.6299 at 5264.978797 found
Include? (y/n):n
The data are plotted with white lines, the model in blue and the model continuum sub-
tracted in red. The green line shows the residuals. The first fit is very bad since we have
not included any background. We will now model the background as a polynomial function.
Enter the following:
cp tmp.config fit system.config
emacs fit system.config &
The file fit system.config is the configuration file used for the model. It included a
header line and the different parameters for each function included in the model (see the
file README fitting tool.txt or Section 7.1.1). We are going to add a new function to
the model, changing the header line from:
0 4 0.2 0.001
to
0 5 0.2 0.001
and adding the description of the new parameters at the end of the file:
0 4 0.2 0.001
poly1d
100 1 -1e12 1e12 -1
0.1 1 -1e12 1e12 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
0 0 0 0 -1
Tutorials: How to handle 3D spectroscopy data 257
Figure 7.12. Example of the use of FIT3D: snapshot of an iteration of fit spec back where a
quasi-stellar object (QSO) spectrum is fitted with three emission lines plus a continuum. The
lines show the input data, overplotting the fitted model. The dot line is the continuum. The
residuals is the line close to zero.
Again, execute the fitting procedure using this new config file as input:
fit spec back.pl spec cen.txt emission lines.txt 4800 5300 none 0 none
fit system.config
The output of the fitting process is illustrated in Figure 7.12. Play with the parameters,
or include a new system to describe the broad line until you get an acceptable fit.
! an ASCII file containing the list of background spectra and the number of them (if
6
If you are not an emacs user or you do not have it installed in your system, please edit
kin back rss.pl and change the preferences to use your favourite editor.
260 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
Figure 7.13. Example of the kinematics analysis. The top-left panel shows the derived model
using the kinematics analysis in Exercise 34. The top-right panel shows the original RSS dataset,
and the bottom-left panel the residuals after subtracting the model.
as illustrated in Figure 7.13. As explained above, it is possible to map the fitted parame-
ters, creating 2D images and interpolating data if required (in case of RSS data). To do
so, run the following:
map interp back.pl sobj.pt.txt out sobj kin.txt out sobj kin 3
1e-12 1
Once the maps are created, it is possible to visualize them using IRAF:
disp out sobj kin flux 00.fits 1
disp out sobj kin vel 00.fits 1
The results are shown in Figure 7.14.
In case of velocity dispersion, it is necessary to subtract, to the observed values, the instru-
mental component. The instrumental component consists of all the dispersions in the lines
that are not physically associated with the studied source but due to the instrument,
night conditions or any other effect. The standard procedure is to perform a kinematic
analysis over a nearby emission line (over the non-sky-subtracted data), and to subtract
the obtained velocity dispersion to the observed object (with a quadratic subtraction).
It is important to note here that the velocity dispersion is fitted assuming a dispersion
in Å, that it is not the proper approach for wide wavelength ranges. Kinematics analyses
are normally performed over a narrow wavelength range, centred on the emission lines of
interest; this is why the current approach is used. It is also easier to decontaminate this
method from instrumental effects in the velocity dispersion.
Tutorials: How to handle 3D spectroscopy data 261
Figure 7.14. Example of the use of FIT3D. It shows the flux intensity map (left panel) together
with a velocity map (right panel) derived with the automatic line fitting performed in Exercise 34.
! when we have low signal-to-noise images of objects, all with the same morphology
interesting:
! when we have a crowded field where we want to deblend the spectra (e.g. the core
in the image (e.g. QSO lenses; Wisotzki et al., 2003);
! when we want to deblend two well-defined components on the same object (e.g. QSO-
of globular clusters or galaxy clusters; Sánchez et al., 2006b); or
galfit3d.pl
This program performs a modelling of an input datacube creating, for each wavelength,
a narrow-band image of the width of a spectral pixel, and fitting it using galfit (Peng
et al., 2002). The output of the fitting is stored in a log file, which consists of the
concatenation of the different outputs (e.g. galfit.01) produced by galfit for each
wavelength. A quality log file is also stored, with an integer flag that indicates if the
fitting process has finished the minimization (1) or ended in abnormal break (0). The
syntax is:
galfit3d.pl INPUT CUBE.fits MOD CUBE.fits RES CUBE.fits
GALFIT CONF LOGFILE QUALITY.txt
GALFIT CONF: A) object.fits, B) output.fits
galfit3d force.pl
This program works like galfit3d.pl, but using as guessed values for the fit the output
values of a previous iteration (INPUT LOGFILE). Each parameter can be smoothed using a
polynomial function, or forced to be a given value. This is indicated with the INPUT NPOLY
! the number of the model in the galfit configuration file whose parameter needs to
configuration file, which is a three-column ASCII file whose elements per line are:
! the order of the polynomial function to fit (if value is an integer) or fixed value (if
value is a float).
The syntax is:
galfit3d force.pl INPUT CUBE.fits MOD CUBE.fits RES CUBE.fits
GALFIT CONF INPUT LOGFILE INPUT NPOLY OUTPUT LOGFILE QUALITY.txt [PLOT]
Figure 7.15. Example of the output of the script imcntr cube.pl: the x (top) and y (bottom)
coordinates of the centroid of a star as a function of the spectral pixel numbers, for a datacube.
To use these scripts, it is first necessary to edit the files and change the line
require /data/sanchez/perl/MY/my.pl;
to fit your own path. It is also necessary to run a mkiraf before running the programs
in the directory where you have the dataset, since these programs are simply IRAF
wrappers.
Exercise 35: Empirical DAR correction (dir: tut3). Using IRAF with DS9 (or
any similar tool), display the channel 300 of a file, e.g. disp star no dar.cube.fits
[*,*,300] and determine the centroid of the star in the field, using imexam or any similar
tool. The centroid is near the coordinates (10.5,11.5). To determine the centroids along
the wavelength, run:
imcntr cube.pl star dar.cube.fits 10.5 11.5 2 2 centers list.txt
10.5 11.5 7.5 14.5
The output of this program is illustrated in Figure 7.15, with the distribution of the x
and y coordinates of the centroid of the star as a function of the spectral pixel numbers
(representing the wavelength). The list of centroids is stored in the centers list.txt
file, used as input in the correction process:
cube shift.pl star dar.cube.fits star no dar.cube.fits
centers.lis 10.5 11.5 shift.cl
The file star no dar.cube.fits contains the DAR corrected datacube.
266 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
Figure 7.16. View of a section of IFS raw data obtained with PMAS in the PPAK mode.
Each dark line corresponds to the projection of a spectrum along the dispersion axis, which, in
this case, corresponds to the X-axis. Spectra are separated by 7 pixels across the Y-axis, and
projected in a few pixels, contaminating the adjacent spectra.
! identification of the position of the spectra on the detector for all pixels along the
pre-processing is complete, the data reduction consists of:
Each of these reduction steps has been treated with a set of command-line routines in
R3D (Sánchez, in preparation), a package specially written for reducing fibre-feed IFS
data. We briefly describe below the different steps required to perform a reduction.
Distortion correction
Most of the spectrographs do not disperse the light homogeneously along the cross-
dispersion axis. The dispersion is distorted, being larger on the edges of the slit than in
the centre. This distortion is sometimes called the ‘C’ distortion, due to its shape on the
CCD. In the case of fibre-feed spectrographs, additional distortions are introduced due
to the way the fibres are placed in the pseudo-slit. Once the spectra are extracted, these
distortions are self-evident. The left panel of Figure 7.17 shows an example of extracted
spectra, corresponding to an arc exposure obtained with PMAS in the PPAK mode. The
‘C’ distortion and the discontinuities in the dispersion solution along the cross-dispersion
may be clearly seen. These discontinuities and the slight differences in the dispersion
from fibre to fibre do not allow a 2D modelling of the distortion map with an analytical
function to be created, neither at the level of the extracted spectra, nor at the level of
the raw data. The distortions have to be corrected fibre to fibre before finding a common
wavelength solution.
R3D performs this correction by a two-step procedure, using arc calibration lamp expo-
sures (like the one in Figure 7.17). First, the peak intensity of a single emission line
is traced along the cross-dispersion axis and shifted to a common reference by a linear
shift. This is done by the program dist cor. This program requires as input the aper-
ture extracted file, the output corrected file, the output first-order distortion file, a flag
indicating whether to smooth (1) or not (0) the solution along the cross-dispersion axis,
the reference pixel in the dispersion axis to look for the emission line peak, the width of
the window in pixels to look for this peak, the number of adjacent pixels to check for the
maximum criterion, and a flag indicating whether to plot the results (1) or not (0). The
syntax is:
dist cor EXTRACTED.fits CORRECTED.fits DISTORTION CORRECTION.txt
SMOOTH[0/1] start index delta index nsearch plot [n fib] [nsigma]
[center index]
After this the intensity peak of a set of selected emission lines is traced, and a poly-
nomial distortion correction is determined in order to recentre all the lines to a common
reference. This is done using the program mdist cor sp. This program requires as input
the output of the first-order correction, an aperture to look for emission lines, a nσ
threshold over the average intensity to look for emission line peaks, the order of the
Tutorials: How to handle 3D spectroscopy data 269
Figure 7.17. Example of extracted spectra from an arc exposure obtained with PMAS in the
PPAK mode, before correcting for the distortion in the dispersion along the cross-dispersion
axis (left panel) and after (right panel). The x-axis corresponds to the dispersion axis.
polynomial function to fit the distortion correction (spectrum by spectrum), the output
file with the distortion-corrected frame, the output file containing the 2D map of the
corrections, the range of pixels in the dispersion direction to consider for the analysis,
and a flag indicating if results will be plotted (1) or not (0). The syntax is:
mdist cor sp EXTRACTED.fits APERTURE NSIGMA NPOLY OUTPUT.fits
DISTORTION.fits NX min NX max plot [NSIGMA2] [BOX] [NY CEN]
Wavelength solution
To derive the wavelength coordinate system, the emission lines of an arc exposure are
first identified using an interactive routine. The distortion-corrected spectra of the arc are
then transformed to a linear wavelength coordinate system by a 1D spline interpolation,
assuming a polynomial transformation between both coordinate systems. The required
transformation is stored in an ASCII file to be applied on the science data at a later
stage. This procedure is done using the disp cor.pl script.
This program requires as input the output of the distortion correction analysis, the
starting wavelength and the wavelength step needed for the final spectral pixels, the
aperture to look for emission lines, the row number of the central spectrum over which
to perform the analysis, width of pixels in the cross-dispersion axis for co-adding the
spectra (to increase the signal-to-noise ratio), the order of the polynomial function to fit
270 Sánchez, Garcı́a-Lorenzo and Pécontal-Rousset
the wavelength solution, the output file containing the spectra transformed to a linear
wavelength coordinate system, the output file containing the dispersion or wavelength
solution applied, and a flag indicating if results will be plotted (1) or not (0). The syntax
is:
disp cor.pl EXTRACTED.fits CRVAL CDELT APERTURE NY SPECTRA
WITDH NPOLY OUTPUT.fits DISPERSION.txt plot
Additional steps
Once wavelength is calibrated, the data need to be corrected from fibre-to-fibre differen-
tial transmission, combined with other exposures (using with off-sets), corrected for the
differential atmospheric refraction, sky-subtracted and flux-calibrated. R3D includes tools
to perform some of these operations, but it is beyond the scope of this quick reference
guide to explain all of them in detail here. Moreover, they may strongly depend on the
instrument used for the acquisition.
Exercise 36: Reducing data with R3D7 (dir: tut6). We are going to perform a
simple data reduction on PMAS Roth et al. (2005) data in the PPAK mode (Kelz et al.,
2006) using R3D. First, we determine the number of spectra and their location along the
cross-dispersion axis for the central column of the CCD:
peak find.pl DomeFF Cont Galaxy.fits 0 2 1 4 2 3.5 0.10
DomeFF Cont Galaxy.peaks
Then, we trace the location of the peak intensity of the spectra along the cross-dispersion
axis of each pixel in the dispersion axis:
trace peaks recursive DomeFF Cont Galaxy.fits 0
DomeFF Cont Galaxy.peaks 2 2 1 4 1 DomeFF Cont Galaxy.trc.fits
Once the location of the spectra in the CCD (trace) has been determined it is possible to
extract them:
extract aper.pl IRAS18131 A.fits 0 DomeFF Cont Galaxy.trc.fits 5
IRAS18131 A.ms.fits
The C distortion correction is then determined and corrected:
dist cor DomeHe ThAr.ms.fits DomeHe ThAr.dc0.fits
DomeHe ThAr.dist.txt 0 662 30 2 1 2 2
mdist cor sp DomeHe ThAr.dc0.fits 3 300 3 DomeHe ThAr.dc1.fits
DomeHe ThAr.dist.fits 17 1011 1 2 5 3
The distortion correction determined using the He lamp exposure can be applied to the
science data:
mdist cor external IRAS18131 A.ms.fits DomeHe ThAr.dist.txt
DomeHe ThAr.dist.fits IRAS18131 A.dc1.fits 0 1.353
Once transformed to a common reference, it is possible to find the wavelength coordinate
solution:
disp cor.pl DomeHe ThAr.dc1.fits 3700 1.6 3 180 3 4
DomeHe ThAr.disp cor.fits DomeHe ThAr.disp.txt 1
7
Data provided by K. Kehring.
Tutorials: How to handle 3D spectroscopy data 271
and apply it to the data:
disp cor external.pl IRAS18131 A.dc1.fits 3700 1.6
IRAS18131 A.disp cor.fits DomeHe ThAr.disp.txt 0
The differences in the fibre-to-fibre transmission can be determined using dome-flat expo-
sures, and correct the science frames from them:
fiber flat.pl mDomeFF Cont IRAS18131 A.disp cor.fits
fiber flat IRAS18131 A.fitsimarith.pl IRAS18131 A.disp cor.fits /
fiber flat IRAS18131 A.fits
IRAS18131 A.fc.fits
REFERENCES
Allington-Smith, J., Content, R. (1998), PASP, 110, 1216
Arribas, S., Mediavilla, E., Watson, F. ET AL. (1998), Fiber Optics in Astronomy III, 152,
149
Arribas, S., Mediavilla, E., Garcı́a-Lorenzo, B., del Burgo, C., Fuensalida, J.J.
(1999), A&AS, 136, 189
Becker, T. (2001), 3D-Spektroskopie hintergrundkontaminierter Einzelobjekte in Galaxien der
lokalen Gruppe, PhD Thesis, Potsdam University
Benn, C., Ellison, S. (1998), La Palma night-sky brightness, ING technical note 115
Fabrika, S., Sholukhova, O., Becker, T., Afanasiev, V., Roth, M., Sánchez, S.F.
(2005), A&A, 437, 217
Filippenko, A.V. (1982), PASP, 94, 715
Garcı́a-Lorenzo, B., Acosta-Pulido, J.A., Megias-Fernández, E. (2002), Galaxies: the
Third Dimension, 282, 501
Garcı́a-Lorenzo, B., Sánchez, S.F., Mediavilla, E. González-Serrano, J.I., Chris-
tensen, L. (2005), ApJ, 621, 146
Glazebrook ET AL. (2001), http://pdl.perl.org/
Kelz, A., Verheijen, M.A.W., Roth, M.M., Bauer, S.M., Becker, T., Paschke, J.,
Popow, E., Sánchez, S.F., Laux, U. (2006), PASP, 118, 129
Kissler-Patig, M., Copin, Y., Ferruit, P., Pécontal-Rousset, A., Roth, M.M. (2004),
AN, 325, 159
Mink, D.J., Kurtz, M.J. (2001), Astronomical Data Analysis Software and Systems X. ASPCS,
238, 491
Pearson, T. (1995), http://www.astro.caltech.edu/∼tjp/pgplot/
Pécontal-Rousset, A., Copin, Y., Ferruit, P. (2004), AN, 325, 163
Peng, C.Y., Ho, L.C., Impey, C.D., Rix, H. (2002), AJ, 124, 266
Roth, M.M., Becker, T., Kelz, A., Schmoll, J. (2004), ApJ, 603, 531
Roth, M.M., Kelz, A., Fechner, T. ET AL. (2005), PASP, 117, 620
Sánchez, S.F. (2004), AN, 325, 167
Sánchez, S.F. (2006), AN, 327, 850
Sánchez, S.F. (2008), Euro3D Visualization Tool: User Guide for E3D, v1.3.7. Research Train-
ing Network. Available online at http://www.aip.de/Euro3D/E3D/#Docu (July 2009)
Sánchez, S.F., Cardiel, N., in preparation
Sánchez S.F., Garcı́a-Lorenzo, B., Mediavilla, E. ET AL. (2004), ApJ, 615, 156
Sánchez S.F., Garcı́a-Lorenzo, B., Jahnke, K. ET AL. (2006a), NewAR, 49, 501
Sánchez, S.F., Cardiel, N., Verheijen, M., Benitez, N. (2006b), Proceedings of the ESO
and Euro3D workshop ‘Science Perspectives for 3D Spectroscopy’, Garching (Germany),
October 10–14 (2005) (astro-ph/0512550)
Walsh, J.R., Roth, M.M. (2002), Msngr, 109, 54
Wisotzki, L., Becker, T., Christensen, L., Helms, A., Jahnke, K., Kelz, A., Roth,
M.M., Sánchez, S.F. (2003), A&A, 408, 455